Spin Locks and Contention. Companion slides for The Art of Multiprocessor Programming by Maurice Herlihy & Nir Shavit

Transcription

1 Spin Locks and Contention Companion slides for The Art of Multiprocessor Programming by Maurice Herlihy & Nir Shavit

2 Focus so far: Correctness and Progress Models Accurate (we never lied to you) But idealized (so we forgot to mention a few things) Protocols Elegant Important But naïve Art of Multiprocessor Programming 2

3 New Focus: Performance Models More complicated (not the same as complex!) Still focus on principles (not soon obsolete) Protocols Elegant (in their fashion) Important (why else would we pay attention) And realistic (your mileage may vary) Art of Multiprocessor Programming 3

4 Kinds of Architectures SISD (Uniprocessor) Single instruction stream Single data stream SIMD (Vector) Single instruction Multiple data MIMD (Multiprocessors) Multiple instruction Multiple data. Art of Multiprocessor Programming 4

5 Kinds of Architectures SISD (Uniprocessor) Single instruction stream Single data stream SIMD (Vector) Single instruction Multiple data MIMD (Multiprocessors) Multiple instruction Multiple data. Our space (1) Art of Multiprocessor Programming 5

6 MIMD Architectures memory Shared Bus Memory Contention Communication Contention Communication Latency Distributed Art of Multiprocessor Programming 6

7 Today: Revisit Mutual Exclusion Performance, not just correctness Proper use of multiprocessor architectures A collection of locking algorithms Art of Multiprocessor Programming 7 (1)

8 What Should you do if you can t get a lock? Keep trying spin or busy-wait Good if delays are short Give up the processor Good if delays are long Always good on uniprocessor Art of Multiprocessor Programming 8 (1)

9 What Should you do if you can t get a lock? Keep trying spin or busy-wait Good if delays are short Give up the processor Good if delays are long Always good on uniprocessor our focus Art of Multiprocessor Programming 9

10 Basic Spin-Lock CS. spin lock critical section Resets lock upon exit Art of Multiprocessor Programming 10

11 Basic Spin-Lock lock introduces sequential bottleneck CS. spin lock critical section Resets lock upon exit Art of Multiprocessor Programming 11

12 Basic Spin-Lock lock suffers from contention CS. spin lock critical section Resets lock upon exit Art of Multiprocessor Programming 12

13 Basic Spin-Lock lock suffers from contention CS. spin lock critical section Resets lock upon exit Notice: these are distinct phenomena Art of Multiprocessor Programming 13

14 Basic Spin-Lock lock suffers from contention CS. spin lock critical section Resets lock upon exit Seq Bottleneck no parallelism Art of Multiprocessor Programming 14

15 Basic Spin-Lock lock suffers from contention CS. spin lock critical section Resets lock upon exit Contention??? Art of Multiprocessor Programming 15

16 Review: Test-and-Set Boolean value Test-and-set (TAS) Swap true with current value Return value tells if prior value was true or false Can reset just by writing false TAS aka getandset Art of Multiprocessor Programming 16

17 Review: Test-and-Set public class AtomicBoolean { boolean value; public synchronized boolean getandset(boolean newvalue) { boolean prior = value; value = newvalue; return prior; } } Art of Multiprocessor Programming 17 (5)

18 Review: Test-and-Set public class AtomicBoolean { boolean value; public synchronized boolean getandset(boolean newvalue) { boolean prior = value; value = newvalue; return prior; } } Package java.util.concurrent.atomic Art of Multiprocessor Programming 18

19 Review: Test-and-Set public class AtomicBoolean { boolean value; public synchronized boolean getandset(boolean newvalue) { boolean prior = value; value = newvalue; return prior; } } Swap old and new values Art of Multiprocessor Programming 19

20 Review: Test-and-Set AtomicBoolean lock = new AtomicBoolean(false) boolean prior = lock.getandset(true) Art of Multiprocessor Programming 20

21 Review: Test-and-Set AtomicBoolean lock = new AtomicBoolean(false) boolean prior = lock.getandset(true) Swapping in true is called test-and-set or TAS Art of Multiprocessor Programming 21 (5)

22 Test-and-Set Locks Locking Lock is free: value is false Lock is taken: value is true Acquire lock by calling TAS If result is false, you win If result is true, you lose Release lock by writing false Art of Multiprocessor Programming 22

23 Test-and-set Lock class TASlock { AtomicBoolean state = new AtomicBoolean(false); void lock() { while (state.getandset(true)) {} } void unlock() { state.set(false); }} Art of Multiprocessor Programming 23

24 Test-and-set Lock class TASlock { AtomicBoolean state = new AtomicBoolean(false); void lock() { while (state.getandset(true)) {} } void unlock() { state.set(false); }} Lock state is AtomicBoolean Art of Multiprocessor Programming 24

25 Test-and-set Lock class TASlock { AtomicBoolean state = new AtomicBoolean(false); void lock() { while (state.getandset(true)) {} } void unlock() { state.set(false); }} Keep trying until lock acquired Art of Multiprocessor Programming 25

26 Test-and-set Lock class TASlock { AtomicBoolean state = new AtomicBoolean(false); Release lock by resetting state to false void lock() { while (state.getandset(true)) {} } void unlock() { state.set(false); }} Art of Multiprocessor Programming 26

27 Space Complexity TAS spin-lock has small footprint N thread spin-lock uses O(1) space As opposed to O(n) Peterson/Bakery How did we overcome the W(n) lower bound? We used a RMW operation Art of Multiprocessor Programming 27

28 Performance Experiment n threads Increment shared counter 1 million times How long should it take? How long does it take? Art of Multiprocessor Programming 28

29 time Graph no speedup because of sequential bottleneck ideal threads Art of Multiprocessor Programming 29

30 time Mystery #1 TAS lock threads Ideal What is going on? Art of Multiprocessor Programming 30

31 Test-and-Test-and-Set Locks Lurking stage Wait until lock looks free Spin while read returns true (lock taken) Pouncing state As soon as lock looks available Read returns false (lock free) Call TAS to acquire lock If TAS loses, back to lurking Art of Multiprocessor Programming 31

32 Test-and-test-and-set Lock class TTASlock { AtomicBoolean state = new AtomicBoolean(false); void lock() { while (true) { while (state.get()) {} if (!state.getandset(true)) return; } } Art of Multiprocessor Programming 32

33 Test-and-test-and-set Lock class TTASlock { AtomicBoolean state = new AtomicBoolean(false); void lock() { while (true) { while (state.get()) {} if (!state.getandset(true)) return; } } Wait until lock looks free Art of Multiprocessor Programming 33

34 Test-and-test-and-set Lock class TTASlock { AtomicBoolean state = new AtomicBoolean(false); void lock() { while (true) { while (state.get()) {} if (!state.getandset(true)) return; } } Then try to acquire it Art of Multiprocessor Programming 34

35 time Mystery #2 TAS lock TTAS lock Ideal threads Art of Multiprocessor Programming 35

36 Mystery Both TAS and TTAS Do the same thing (in our model) Except that TTAS performs much better than TAS Neither approaches ideal Art of Multiprocessor Programming 36

37 Opinion Our memory abstraction is broken TAS & TTAS methods Are provably the same (in our model) Except they aren t (in field tests) Need a more detailed model Art of Multiprocessor Programming 37

38 Bus-Based Architectures cache cache Bus cache memory Art of Multiprocessor Programming 38

39 Bus-Based Architectures Random access memory (10s of cycles) cache cache Bus cache memory Art of Multiprocessor Programming 39

40 Bus-Based Architectures Shared Bus Broadcast medium One broadcaster at a time Processors and memory all snoop cache cache Bus cache memory Art of Multiprocessor Programming 40

41 Per-Processor Caches Small Fast: 1 or 2 cycles Address & state information Bus-Based Architectures cache cache Bus cache memory Art of Multiprocessor Programming 41

42 Granularity Caches operate at a larger granularity than a word Cache line: fixed-size block containing the address (today 64 or 128 bytes) Art of Multiprocessor Programming 42

43 Locality If you use an address now, you will probably use it again soon Fetch from cache, not memory If you use an address now, you will probably use a nearby address soon In the same cache line Art of Multiprocessor Programming 43

44 L1 and L2 Caches L2 L1 Art of Multiprocessor Programming 44

45 L1 and L2 Caches L2 L1 Small & fast 1 or 2 cycles Art of Multiprocessor Programming 45

46 Larger and slower 10s of cycles ~128 byte line L1 and L2 Caches L2 L1 Art of Multiprocessor Programming 46

47 Jargon Watch Cache hit I found what I wanted in my cache Good Thing Art of Multiprocessor Programming 47

48 Jargon Watch Cache hit I found what I wanted in my cache Good Thing Cache miss I had to shlep all the way to memory for that data Bad Thing Art of Multiprocessor Programming 48

49 Cave Canem This model is still a simplification But not in any essential way Illustrates basic principles Will discuss complexities later Art of Multiprocessor Programming 49

50 When a Cache Becomes Full Need to make room for new entry By evicting an existing entry Need a replacement policy Usually some kind of least recently used heuristic Art of Multiprocessor Programming 50

51 Fully Associative Cache Any line can be anywhere in the cache Advantage: can replace any line Disadvantage: hard to find lines Art of Multiprocessor Programming 51

52 Direct Mapped Cache Every address has exactly 1 slot Advantage: easy to find a line Disadvantage: must replace fixed line Art of Multiprocessor Programming 52

53 K-way Set Associative Cache Each slot holds k lines Advantage: pretty easy to find a line Advantage: some choice in replacing line Art of Multiprocessor Programming 53

54 Multicore Set Associativity k is 8 or even 16 and growing Why? Because cores share sets Threads cut effective size if accessing different data Art of Multiprocessor Programming 54

55 Cache Coherence A and B both cache address x A writes to x Updates cache How does B find out? Many cache coherence protocols in literature Art of Multiprocessor Programming 55

56 MESI Modified Have modified cached data, must write back to memory Art of Multiprocessor Programming 56

57 MESI Modified Have modified cached data, must write back to memory Exclusive Not modified, I have only copy Art of Multiprocessor Programming 57

58 MESI Modified Have modified cached data, must write back to memory Exclusive Not modified, I have only copy Shared Not modified, may be cached elsewhere Art of Multiprocessor Programming 58

59 MESI Modified Have modified cached data, must write back to memory Exclusive Not modified, I have only copy Shared Not modified, may be cached elsewhere Invalid Cache contents not meaningful Art of Multiprocessor Programming 59

60 Processor Issues Load Request load x cache cache Bus cache memory data Art of Multiprocessor Programming 60

61 Memory Responds E cache cache cache Bus Bus Got it! memory data Art of Multiprocessor Programming 61

62 Processor Issues Load Request Load x E data cache cache Bus memory data Art of Multiprocessor Programming 62

63 Other Processor Responds Got it ES data S cache cache Bus Bus memory data Art of Multiprocessor Programming 63

64 Modify Cached Data S data S data cache Bus memory data Art of Multiprocessor Programming 64

65 Write-Through Cache Write x! S data S data cache Bus memory data Art of Multiprocessor Programming 65

66 Write-Through Caches Immediately broadcast changes Good Memory, caches always agree More read hits, maybe Bad Bus traffic on all writes Most writes to unshared data For example, loop indexes Art of Multiprocessor Programming 66

67 Write-Through Caches Immediately broadcast changes Good Memory, caches always agree More read hits, maybe Bad Bus traffic on all writes Most writes to unshared data For example, loop indexes show stoppers Art of Multiprocessor Programming 67

68 Write-Back Caches Accumulate changes in cache Write back when line evicted Need the cache for something else Another processor wants it Art of Multiprocessor Programming 68

69 Invalidate Invalidate x SI cache data MS data cache Bus memory data Art of Multiprocessor Programming 69

70 Invalidate cache data Bus cache This cache acquires write permission memory data Art of Multiprocessor Programming 70

71 Invalidate Other caches lose read permission cache data Bus cache This cache acquires write permission memory data Art of Multiprocessor Programming 71

72 Invalidate Memory provides data only if not present in any cache, so no need to change it now (expensive) cache data Bus cache memory data Art of Multiprocessor Programming 72

73 Mutual Exclusion What do we want to optimize? Bus bandwidth used by spinning threads Release/Acquire latency Acquire latency for idle lock Art of Multiprocessor Programming 73

74 Simple TASLock TAS invalidates cache lines Spinners Miss in cache Go to bus Thread wants to release lock delayed behind spinners Art of Multiprocessor Programming 74

75 Test-and-test-and-set Wait until lock looks free Spin on local cache No bus use while lock busy Problem: when lock is released Invalidation storm Art of Multiprocessor Programming 75

76 Local Spinning while Lock is Busy busy busy Bus busy memory busy Art of Multiprocessor Programming 76

77 On Release invalid invalid Bus free memory free Art of Multiprocessor Programming 77

78 Everyone misses, rereads On Release miss invalid invalid miss Bus free memory free Art of Multiprocessor Programming 78 (1)

79 Everyone tries TAS On Release TAS( ) invalid TAS( ) invalid Bus free memory free Art of Multiprocessor Programming 79 (1)

80 Problems Everyone misses Reads satisfied sequentially Everyone does TAS Invalidates others caches Eventually quiesces after lock acquired How long does this take? Art of Multiprocessor Programming 80

81 Measuring Quiescence Time Acquire lock Pause without using bus Use bus heavily P 1 P 2 P n If pause > quiescence time, critical section duration independent of number of threads If pause < quiescence time, critical section duration slower with more threads Art of Multiprocessor Programming 81

82 time Quiescence Time Increses linearly with the number of processors for bus architecture threads Art of Multiprocessor Programming 82

83 time Mystery Explained TAS lock TTAS lock Ideal threads Better than TAS but still not as good as ideal Art of Multiprocessor Programming 83

84 Solution: Introduce Delay If the lock looks free But I fail to get it There must be contention Better to back off than to collide again time r 2 d r 1 d d spin lock Art of Multiprocessor Programming 84

85 Dynamic Example: Exponential Backoff time 4d 2d d spin lock If I fail to get lock Wait random duration before retry Each subsequent failure doubles expected wait Art of Multiprocessor Programming 85

86 Exponential Backoff Lock public class Backoff implements lock { public void lock() { int delay = MIN_DELAY; while (true) { while (state.get()) {} if (!lock.getandset(true)) return; sleep(random() % delay); if (delay < MAX_DELAY) delay = 2 * delay; }}} Art of Multiprocessor Programming 86

87 Exponential Backoff Lock public class Backoff implements lock { public void lock() { int delay = MIN_DELAY; while (true) { while (state.get()) {} if (!lock.getandset(true)) return; sleep(random() % delay); if (delay < MAX_DELAY) delay = 2 * delay; }}} Fix minimum delay Art of Multiprocessor Programming 87

88 Exponential Backoff Lock public class Backoff implements lock { public void lock() { int delay = MIN_DELAY; while (true) { while (state.get()) {} if (!lock.getandset(true)) return; sleep(random() % delay); if (delay < MAX_DELAY) delay = 2 * delay; }}} Wait until lock looks free Art of Multiprocessor Programming 88

89 Exponential Backoff Lock public class Backoff implements lock { public void lock() { int delay = MIN_DELAY; while (true) { while (state.get()) {} if (!lock.getandset(true)) return; sleep(random() % delay); if (delay < MAX_DELAY) delay = 2 * delay; }}} If we win, return Art of Multiprocessor Programming 89

90 Exponential Backoff Lock public class Backoff implements lock { public void lock() { int delay = MIN_DELAY; while (true) { while (state.get()) {} if (!lock.getandset(true)) return; sleep(random() % delay); if (delay < MAX_DELAY) delay = 2 * delay; }}} Back off for random duration Art of Multiprocessor Programming 90

91 Exponential Backoff Lock public class Backoff implements lock { public void lock() { int delay = MIN_DELAY; while (true) { while (state.get()) {} if (!lock.getandset(true)) return; sleep(random() % delay); if (delay < MAX_DELAY) delay = 2 * delay; }}} Double max delay, within reason Art of Multiprocessor Programming 91

92 time Spin-Waiting Overhead TTAS Lock Backoff lock threads Art of Multiprocessor Programming 92

93 Backoff: Other Issues Good Easy to implement Beats TTAS lock Bad Must choose parameters carefully Not portable across platforms Art of Multiprocessor Programming 93

94 Idea Avoid useless invalidations By keeping a queue of threads Each thread Notifies next in line Without bothering the others Art of Multiprocessor Programming 95

95 Anderson Queue Lock next idle flags T F F F F F F F Art of Multiprocessor Programming 96

96 Anderson Queue Lock next acquiring getandincrement flags T F F F F F F F Art of Multiprocessor Programming 97

97 Anderson Queue Lock next acquiring getandincrement flags T F F F F F F F Art of Multiprocessor Programming 98

98 Anderson Queue Lock next acquired Mine! flags T F F F F F F F Art of Multiprocessor Programming 99

99 Anderson Queue Lock next acquired acquiring flags T F F F F F F F Art of Multiprocessor Programming 100

100 Anderson Queue Lock next acquired acquiring flags getandincrement T F F F F F F F Art of Multiprocessor Programming 101

101 Anderson Queue Lock next acquired acquiring flags getandincrement T F F F F F F F Art of Multiprocessor Programming 102

102 Anderson Queue Lock next acquired acquiring flags T F F F F F F F Art of Multiprocessor Programming 103

103 Anderson Queue Lock next released acquired flags T T F F F F F F Art of Multiprocessor Programming 104

104 Anderson Queue Lock next released acquired flags Yow! T T F F F F F F Art of Multiprocessor Programming 105

105 Anderson Queue Lock class ALock implements Lock { boolean[] flags={true,false,,false}; AtomicInteger next = new AtomicInteger(0); ThreadLocal<Integer> myslot; Art of Multiprocessor Programming 106

106 Anderson Queue Lock class ALock implements Lock { boolean[] flags={true,false,,false}; AtomicInteger next = new AtomicInteger(0); ThreadLocal<Integer> myslot; One flag per thread Art of Multiprocessor Programming 107

107 Anderson Queue Lock class ALock implements Lock { boolean[] flags={true,false,,false}; AtomicInteger next = new AtomicInteger(0); ThreadLocal<Integer> myslot; Next flag to use Art of Multiprocessor Programming 108

108 Anderson Queue Lock class ALock implements Lock { boolean[] flags={true,false,,false}; AtomicInteger next = new AtomicInteger(0); ThreadLocal<Integer> myslot; Thread-local variable Art of Multiprocessor Programming 109

109 Anderson Queue Lock public lock() { myslot = next.getandincrement(); while (!flags[myslot % n]) {}; flags[myslot % n] = false; } public unlock() { flags[(myslot+1) % n] = true; } Art of Multiprocessor Programming 110

110 Anderson Queue Lock public lock() { myslot = next.getandincrement(); while (!flags[myslot % n]) {}; flags[myslot % n] = false; } public unlock() { flags[(myslot+1) % n] = true; } Take next slot Art of Multiprocessor Programming 111

111 Anderson Queue Lock public lock() { myslot = next.getandincrement(); while (!flags[myslot % n]) {}; flags[myslot % n] = false; } public unlock() { flags[(myslot+1) % n] = true; } Spin until told to go Art of Multiprocessor Programming 112

112 Anderson Queue Lock public lock() { myslot = next.getandincrement(); while (!flags[myslot % n]) {}; flags[myslot % n] = false; } public unlock() { flags[(myslot+1) % n] = true; } Prepare slot for re-use Art of Multiprocessor Programming 113

113 Anderson Queue Lock public lock() { Tell next thread to go myslot = next.getandincrement(); while (!flags[myslot % n]) {}; flags[myslot % n] = false; } public unlock() { flags[(myslot+1) % n] = true; } Art of Multiprocessor Programming 114

114 Local Spinning next flags released acquired Spin on my bit T F F F F F F F Unfortunately many bits share cache line Art of Multiprocessor Programming 115

115 next Result: contention flags False Sharing released acquired T F F F F F F F Line 1Art of Multiprocessor Programming Line 2 Spin on my Spinning thread bit gets cache invalidation on account of store by threads it is not waiting for 116

116 The Solution: Padding next flags released acquired Spin on my line T / / / F / / / Line 1Art of Multiprocessor Programming Line 2 117

117 Performance TTAS queue Shorter handover than backoff Curve is practically flat Scalable performance Art of Multiprocessor Programming 118

118 Anderson Queue Lock Good First truly scalable lock Simple, easy to implement Back to FCFS order (like Bakery) Art of Multiprocessor Programming 119

119 Bad Anderson Queue Lock Space hog One bit per thread one cache line per thread What if unknown number of threads? What if small number of actual contenders? Art of Multiprocessor Programming 120

120 CLH Lock FCFS order Small, constant-size overhead per thread Art of Multiprocessor Programming 121

121 Initially idle tail false Art of Multiprocessor Programming 122

122 Initially idle tail false Queue tail Art of Multiprocessor Programming 123

123 Initially idle tail false Lock is free Art of Multiprocessor Programming 124

124 Initially idle tail false Art of Multiprocessor Programming 125

125 Purple Wants the Lock acquiring tail false Art of Multiprocessor Programming 126

126 Purple Wants the Lock acquiring tail false true Art of Multiprocessor Programming 127

127 Purple Wants the Lock acquiring Swap tail false true Art of Multiprocessor Programming 128

128 Purple Has the Lock acquired tail false true Art of Multiprocessor Programming 129

129 Red Wants the Lock acquired acquiring tail false true true Art of Multiprocessor Programming 130

130 Red Wants the Lock acquired acquiring Swap tail false true true Art of Multiprocessor Programming 131

131 Red Wants the Lock acquired acquiring tail false true true Art of Multiprocessor Programming 132

132 Red Wants the Lock acquired acquiring tail false true true Art of Multiprocessor Programming 133

133 Red Wants the Lock acquired acquiring Implicit Linked list tail false true true Art of Multiprocessor Programming 134

134 Red Wants the Lock acquired acquiring tail false true true Art of Multiprocessor Programming 135

135 Red Wants the Lock acquired acquiring tail false true true true Actually, it spins on cached copy Art of Multiprocessor Programming 136

136 Purple Releases release acquiring false Bingo! tail false false true Art of Multiprocessor Programming 137

137 Purple Releases released acquired tail true Art of Multiprocessor Programming 138

138 Space Usage Let L = number of locks N = number of threads ALock O(LN) CLH lock O(L+N) Art of Multiprocessor Programming 139

139 CLH Queue Lock class QNode { AtomicBoolean locked = new AtomicBoolean(true); } Art of Multiprocessor Programming 140

140 CLH Queue Lock class QNode { AtomicBoolean locked = new AtomicBoolean(true); } Not released yet Art of Multiprocessor Programming 141

141 CLH Queue Lock class CLHLock implements Lock { AtomicReference<QNode> tail; ThreadLocal<QNode> mynode = new QNode(); public void lock() { QNode pred = tail.getandset(mynode); while (pred.locked) {} }} Art of Multiprocessor Programming 142

142 CLH Queue Lock class CLHLock implements Lock { AtomicReference<QNode> tail; ThreadLocal<QNode> mynode = new QNode(); public void lock() { QNode pred = tail.getandset(mynode); while (pred.locked) {} }} Queue tail Art of Multiprocessor Programming 143

143 CLH Queue Lock class CLHLock implements Lock { AtomicReference<QNode> tail; ThreadLocal<QNode> mynode = new QNode(); public void lock() { QNode pred = tail.getandset(mynode); while (pred.locked) {} }} Thread-local QNode Art of Multiprocessor Programming 144

144 CLH Queue Lock class CLHLock implements Lock { AtomicReference<QNode> tail; ThreadLocal<QNode> mynode = new QNode(); public void lock() { QNode pred = tail.getandset(mynode); while (pred.locked) {} }} Swap in my node Art of Multiprocessor Programming 145

145 CLH Queue Lock class CLHLock implements Lock { AtomicReference<QNode> tail; ThreadLocal<QNode> mynode = new QNode(); public void lock() { QNode pred = tail.getandset(mynode); while (pred.locked) {} }} Spin until predecessor releases lock Art of Multiprocessor Programming 146

146 CLH Queue Lock Class CLHLock implements Lock { public void unlock() { mynode.locked.set(false); mynode = pred; } } Art of Multiprocessor Programming 147

147 CLH Queue Lock Class CLHLock implements Lock { public void unlock() { mynode.locked.set(false); mynode = pred; } } Notify successor Art of Multiprocessor Programming 148

148 CLH Queue Lock Class CLHLock implements Lock { } public void unlock() { } mynode.locked.set(false); mynode = pred; Recycle predecessor s node Art of Multiprocessor Programming 149

149 CLH Queue Lock Class CLHLock implements Lock { } public void unlock() { } mynode.locked.set(false); mynode = pred; (we don t actually reuse mynode. Code in book shows how it s done.) Art of Multiprocessor Programming 150

150 CLH Lock Good Lock release affects predecessor only Small, constant-sized space Bad Doesn t work for uncached NUMA architectures Art of Multiprocessor Programming 151

151 NUMA and cc-numa Architectures Acronym: Non-Uniform Memory Architecture ccnuma = cache coherent NUMA Illusion: Flat shared memory Truth: No caches (sometimes) Some memory regions faster than others Art of Multiprocessor Programming 152

152 NUMA Machines Spinning on local memory is fast Art of Multiprocessor Programming 153

153 NUMA Machines Spinning on remote memory is slow Art of Multiprocessor Programming 154

154 CLH Lock Each thread spins on predecessor s memory Could be far away Art of Multiprocessor Programming 155

155 MCS Lock FCFS order Spin on local memory only Small, Constant-size overhead Art of Multiprocessor Programming 156

156 Initially idle tail false Art of Multiprocessor Programming 157

157 Acquiring acquiring (allocate QNode) tail false true Art of Multiprocessor Programming 158

158 Acquiring acquired tail swap false true Art of Multiprocessor Programming 159

159 Acquiring acquired tail false true Art of Multiprocessor Programming 160

160 Acquired acquired tail false true Art of Multiprocessor Programming 161

161 Acquiring acquired acquiring tail swap false true Art of Multiprocessor Programming 162

162 Acquiring acquired acquiring tail false true Art of Multiprocessor Programming 163

163 Acquiring acquired acquiring tail false true Art of Multiprocessor Programming 164

164 Acquiring acquired acquiring tail false true Art of Multiprocessor Programming 165

165 Acquiring acquired acquiring tail true false true Art of Multiprocessor Programming 166

166 Acquiring acquired acquiring tail true Yes! false true Art of Multiprocessor Programming 167

167 MCS Queue Lock class QNode { volatile boolean locked = false; volatile qnode next = null; } Art of Multiprocessor Programming 168

168 MCS Queue Lock class MCSLock implements Lock { AtomicReference tail; public void lock() { QNode qnode = new QNode(); QNode pred = tail.getandset(qnode); if (pred!= null) { qnode.locked = true; pred.next = qnode; while (qnode.locked) {} }}} Art of Multiprocessor Programming 169

169 MCS Queue Lock class MCSLock implements Lock { AtomicReference tail; public void lock() { QNode qnode = new QNode(); QNode pred = tail.getandset(qnode); if (pred!= null) { qnode.locked = true; pred.next = qnode; while (qnode.locked) {} }}} Make a QNode Art of Multiprocessor Programming 170

170 MCS Queue Lock class MCSLock implements Lock { AtomicReference tail; public void lock() { QNode qnode = new QNode(); QNode pred = tail.getandset(qnode); if (pred!= null) { qnode.locked = true; pred.next = qnode; while (qnode.locked) {} }}} add my Node to the tail of queue Art of Multiprocessor Programming 171

171 MCS Queue Lock class MCSLock implements Lock { AtomicReference tail; public void lock() { QNode qnode = new QNode(); QNode pred = tail.getandset(qnode); if (pred!= null) { qnode.locked = true; pred.next = qnode; while (qnode.locked) {} }}} Fix if queue was non-empty Art of Multiprocessor Programming 172

172 MCS Queue Lock class MCSLock implements Lock { AtomicReference tail; public void lock() { QNode qnode = new QNode(); Wait until unlocked QNode pred = tail.getandset(qnode); if (pred!= null) { qnode.locked = true; pred.next = qnode; while (qnode.locked) {} }}} Art of Multiprocessor Programming 173

173 Purple Release releasing swap false false Art of Multiprocessor Programming 174

174 Purple Release releasing I don t see a successor. But by looking at the queue, I see another thread is active swap false false Art of Multiprocessor Programming 175

175 Purple Release releasing I don t see a successor. But by looking at the queue, I see another thread is active swap false false I have to release that thread so must wait for it to identify its node Art of Multiprocessor Programming 176

176 Purple Release releasing prepare to spin true false Art of Multiprocessor Programming 177

177 Purple Release releasing spinning true false Art of Multiprocessor Programming 178

178 Purple Release releasing spinning false true false Art of Multiprocessor Programming 179

179 Purple Release releasing Acquired lock false true false Art of Multiprocessor Programming 180

180 MCS Queue Unlock class MCSLock implements Lock { AtomicReference tail; public void unlock() { if (qnode.next == null) { if (tail.cas(qnode, null) return; while (qnode.next == null) {} } qnode.next.locked = false; }} Art of Multiprocessor Programming 181

181 MCS Queue Lock class MCSLock implements Lock { AtomicReference tail; public void unlock() { if (qnode.next == null) { } if (tail.cas(qnode, null) return; while (qnode.next == null) {} qnode.next.locked = false; }} Missing successor? Art of Multiprocessor Programming 182

182 MCS Queue Lock class MCSLock implements Lock { If really no successor, return AtomicReference tail; public void unlock() { if (qnode.next == null) { if (tail.cas(qnode, null) return; while (qnode.next == null) {} } qnode.next.locked = false; }} Art of Multiprocessor Programming 183

183 MCS Queue Lock class MCSLock implements Lock { Otherwise wait for successor to catch up AtomicReference tail; public void unlock() { if (qnode.next == null) { if (tail.cas(qnode, null) return; while (qnode.next == null) {} } qnode.next.locked = false; }} Art of Multiprocessor Programming 184

184 MCS Queue Lock class MCSLock implements Lock { AtomicReference queue; Pass lock to successor public void unlock() { if (qnode.next == null) { if (tail.cas(qnode, null) return; while (qnode.next == null) {} } qnode.next.locked = false; }} Art of Multiprocessor Programming 185

185 Abortable Locks What if you want to give up waiting for a lock? For example Timeout Database transaction aborted by user Art of Multiprocessor Programming 186

186 Back-off Lock Aborting is trivial Just return from lock() call Extra benefit: No cleaning up Wait-free Immediate return Art of Multiprocessor Programming 187

187 Queue Locks Can t just quit Thread in line behind will starve Need a graceful way out Art of Multiprocessor Programming 188

188 Queue Locks spinning spinning spinning true true true Art of Multiprocessor Programming 189

189 Queue Locks locked spinning spinning false true true Art of Multiprocessor Programming 190

190 Queue Locks locked spinning false true Art of Multiprocessor Programming 191

191 Queue Locks locked false Art of Multiprocessor Programming 192

192 Queue Locks spinning spinning spinning true true true Art of Multiprocessor Programming 193

193 Queue Locks spinning spinning true true true Art of Multiprocessor Programming 194

194 Queue Locks locked spinning false true true Art of Multiprocessor Programming 195

195 Queue Locks spinning false true Art of Multiprocessor Programming 196

196 Queue Locks pwned false true Art of Multiprocessor Programming 197

197 Abortable CLH Lock When a thread gives up Removing node in a wait-free way is hard Idea: let successor deal with it. Art of Multiprocessor Programming 198

198 Initially idle Pointer to predecessor (or null) tail A Art of Multiprocessor Programming 199

199 Initially idle Distinguished available node means lock is free tail A Art of Multiprocessor Programming 200

200 acquiring Acquiring tail A Art of Multiprocessor Programming 201

201 acquiring Acquiring Null predecessor means lock not released or aborted A Art of Multiprocessor Programming 202

202 acquiring Acquiring Swap A Art of Multiprocessor Programming 203

203 acquiring Acquiring A Art of Multiprocessor Programming 204

204 Acquired locked Reference to AVAILABLE means lock is free. A Art of Multiprocessor Programming 205

205 Normal Case locked spinning spinning Null means lock is not free & request not aborted Art of Multiprocessor Programming 206

206 One Thread Aborts locked Timed out spinning Art of Multiprocessor Programming 207

207 Successor Notices locked Timed out spinning Non-Null means predecessor aborted Art of Multiprocessor Programming 208

208 Recycle Predecessor s Node locked spinning Art of Multiprocessor Programming 209

209 Spin on Earlier Node locked spinning Art of Multiprocessor Programming 210

210 Spin on Earlier Node released spinning A The lock is now mine Art of Multiprocessor Programming 211

211 Time-out Lock public class TOLock implements Lock { static QNode AVAILABLE = new QNode(); AtomicReference<QNode> tail; ThreadLocal<QNode> mynode; Art of Multiprocessor Programming 212

212 Time-out Lock public class TOLock implements Lock { static QNode AVAILABLE = new QNode(); AtomicReference<QNode> tail; ThreadLocal<QNode> mynode; AVAILABLE node signifies free lock Art of Multiprocessor Programming 213

213 Time-out Lock public class TOLock implements Lock { static QNode AVAILABLE = new QNode(); AtomicReference<QNode> tail; ThreadLocal<QNode> mynode; Tail of the queue Art of Multiprocessor Programming 214

214 Time-out Lock public class TOLock implements Lock { static QNode AVAILABLE = new QNode(); AtomicReference<QNode> tail; ThreadLocal<QNode> mynode; Remember my node Art of Multiprocessor Programming 215

215 Time-out Lock public boolean lock(long timeout) { QNode qnode = new QNode(); mynode.set(qnode); qnode.prev = null; QNode mypred = tail.getandset(qnode); if (mypred== null mypred.prev == AVAILABLE) { return true; } Art of Multiprocessor Programming 216

216 Time-out Lock public boolean lock(long timeout) { QNode qnode = new QNode(); mynode.set(qnode); qnode.prev = null; QNode mypred = tail.getandset(qnode); if (mypred == null mypred.prev == AVAILABLE) { return true; } Create & initialize node Art of Multiprocessor Programming 217

217 Time-out Lock public boolean lock(long timeout) { QNode qnode = new QNode(); mynode.set(qnode); qnode.prev = null; QNode mypred = tail.getandset(qnode); if (mypred == null mypred.prev == AVAILABLE) { return true; } Swap with tail Art of Multiprocessor Programming 218

218 Time-out Lock public boolean lock(long timeout) { QNode qnode = new QNode(); mynode.set(qnode); qnode.prev = null; QNode mypred = tail.getandset(qnode); if (mypred == null }... mypred.prev == AVAILABLE) { return true; If predecessor absent or released, we are done Art of Multiprocessor Programming 219

219 locked Time-out Lock spinning spinning long start = now(); while (now()- start < timeout) { QNode predpred = mypred.prev; if (predpred == AVAILABLE) { return true; } else if (predpred!= null) { mypred = predpred; } } Art of Multiprocessor Programming 220

220 Time-out Lock long start = now(); while (now()- start < timeout) { } QNode predpred = mypred.prev; if (predpred == AVAILABLE) { return true; } else if (predpred!= null) { } mypred = predpred; Keep trying for a while Art of Multiprocessor Programming 221

221 Time-out Lock long start = now(); while (now()- start < timeout) { } QNode predpred = mypred.prev; if (predpred == AVAILABLE) { return true; } else if (predpred!= null) { } mypred = predpred; Spin on predecessor s prev field Art of Multiprocessor Programming 222

222 Time-out Lock long start = now(); while (now()- start < timeout) { QNode predpred = mypred.prev; if (predpred == AVAILABLE) { return true; } else if (predpred!= null) { mypred = predpred; } } Predecessor released lock Art of Multiprocessor Programming 223

223 Time-out Lock long start = now(); while (now()- start < timeout) { } QNode predpred = mypred.prev; if (predpred == AVAILABLE) { return true; } else if (predpred!= null) { } mypred = predpred; Predecessor aborted, advance one Art of Multiprocessor Programming 224

224 Time-out Lock if (!tail.compareandset(qnode, mypred)) qnode.prev = mypred; return false; } } What do I do when I time out? Art of Multiprocessor Programming 225

225 Time-out Lock if (!tail.compareandset(qnode, mypred)) qnode.prev = mypred; return false; } } Do I have a successor? If CAS fails, I do. Tell it about mypred Art of Multiprocessor Programming 226

226 Time-out Lock if (!tail.compareandset(qnode, mypred)) qnode.prev = mypred; return false; } } If CAS succeeds: no successor, simply return false Art of Multiprocessor Programming 227

227 Time-Out Unlock public void unlock() { QNode qnode = mynode.get(); if (!tail.compareandset(qnode, null)) qnode.prev = AVAILABLE; } Art of Multiprocessor Programming 228

228 Time-out Unlock public void unlock() { QNode qnode = mynode.get(); if (!tail.compareandset(qnode, null)) qnode.prev = AVAILABLE; } If CAS failed: successor exists, notify it can enter Art of Multiprocessor Programming 229

229 Timing-out Lock public void unlock() { QNode qnode = mynode.get(); if (!tail.compareandset(qnode, null)) qnode.prev = AVAILABLE; } CAS successful: set tail to null, no clean up since no successor waiting Art of Multiprocessor Programming 230

230 Fairness and NUMA Locks MCS lock mechanics are aware of NUMA Lock Fairness is FCFS Is this a good fit with NUMA and Cache-Coherent NUMA machines?

231 Lock Data Access in NUMA Machine Node 1 CS MCS lock various memory locations Node 2

232 Who s the Unfairest of Them All? locality crucial to NUMA performance Big gains if threads from same node/cluster obtain lock consecutively Unfairness pays

233 Hierarchical Backoff Lock (HBO) Back off less for thread from same node time 4d 2d d CS Unfairness is key to performance Global T&T&S lock time 4d 2d d

234 Hierarchical Backoff Lock (HBO) Advantages: Simple, improves locality Disadvantages: Requires platform specific tuning Unstable Unfair Continuous invalidations on shared global lock word

235 Hierarchical CLH Lock (HCLH) Each thread spins on cached copy of predecessor s node Local Tail Thread at local head splices local queue into global queue Local CLH queue Global Tail CAS() CAS() Local Tail Local CLH queue CS CAS()

236 Hierarchical CLH Lock (HCLH) HCLH HBO Threads access 4 cache lines in CS

237 Hierarchical CLH Lock (HCLH) Advantages: Improved locality Local spinning Fair Disadvantages: Complex code implies long common path Splicing into both local and global requires CAS Hard to get long local sequences

238

239 Lock Cohorting General technique for converting almost any lock into a NUMA lock Allows combining different lock types But need these locks to have certain properties (will discuss shortly)

240 Lock Cohorting Non-empty cohort empty cohort Acquire local lock and proceed to critical section Local Lock On release: if nonempty cohort of waiting threads, release only local lock; leave mark Thread that acquired local lock can now acquire global lock Local Lock Global Lock CS On release: since cohort is empty must release global lock to avoid deadlock

241 Thread Obliviousness A lock is thread-oblivious if After being acquired by one thread, Can be released by another Art of Multiprocessor Programming 242

242 Cohort Detection A lock x provides cohort detection if It can tell whether any thread is trying to acquire it Art of Multiprocessor Programming 243

243 Lock Cohorting Two levels of locking Global lock: thread oblivious Thread acquiring the lock can be different than one releasing it Local lock: cohort detection Thread releasing can detect if some thread is waiting to acquire it

244 Two new states: acquire local and acquire global. Do we own global lock? Lock Cohorting: C-BO-MCS In Lock, cohortlock Local MCS lock tail detection by checking successor pointer CAS() False False True Global backoff lock Bound number of Local consecutive MCS lock acquires to control tail unfairness CAS() False False True time 4d 2d d BO Lock is thread oblivious by definition CS

245 How to add cohort detection Lock Cohorting: property to BO lock? C-BO-BO Lock 4d 2d d Global backoff lock CS time 4d 2d d 4d 2d d As noted BO Lock is thread oblivious

246 Add successorexists Lock Cohorting: C-BO-BO Lock field before attempting to acquire local lock. successorexists reset on lock release. 4d 2d d Release might overwrite another successor s Global write but we don t backoff care why? lock CS time 4d 2d d 4d 2d d

247 C-BO-BO Aborting thread is a resets Time-Out successorexists field before leaving local lock. Spinning threads set it to true. NUMA Lock 4d 2d d BO locks trivially abortable Global backoff lock If releasing thread finds successorexists time false, 4dit releases global lock 2d d CS 4d 2d d

248 Lock Cohorting Advantages: Great locality Low contention on shared lock Practically no tuning Has whatever properties you want: Can be more or less fair, abortable just choose the appropriate type of locks Disadvantages: Must tune fairness parameters

249 Lock Cohorting C-BO-MCS C-BO-BO HCLH HBO

250 Throughput in CR-nCRs per sec Time-Out (Abortable) Lock Cohorting A-BO-CLH (time-out lock + BO) 4e e+06 3e e+06 A-BO-BO a-clh a-hbo a-bo-bo a-bo-clh (CAS) a-bo-clh 2e e+06 1e Abortable CLH (our time-out lock) and HBO Number of Threads

251 One Lock To Rule Them All? TTAS+Backoff, CLH, MCS, ToLock Each better than others in some way There is no one solution Lock we pick really depends on: the application the hardware which properties are important Art of Multiprocessor Programming 253

252 This work is licensed under a Creative Commons Attribution- ShareAlike 2.5 License. You are free: to Share to copy, distribute and transmit the work to Remix to adapt the work Under the following conditions: Attribution. You must attribute the work to The Art of Multiprocessor Programming (but not in any way that suggests that the authors endorse you or your use of the work). Share Alike. If you alter, transform, or build upon this work, you may distribute the resulting work only under the same, similar or a compatible license. For any reuse or distribution, you must make clear to others the license terms of this work. The best way to do this is with a link to Any of the above conditions can be waived if you get permission from the copyright holder. Nothing in this license impairs or restricts the author's moral rights. Art of Multiprocessor Programming 254