Racing To Win: Using Race Conditions to Build Correct and Concurrent Software
•
0 recomendaciones•1,122 vistas
If you've ever worked on parallel or multiprocessor software, you've almost certainly encountered bugs owning to race conditions between concurrently-executing components. While race conditions intuitively seem bad, it turns out there are cases in which we can use them to our advantage! In this talk, we'll discuss a number of ways that race conditions are used in improving throughput and reducing latency in high-performance systems, without sacrificing correctness along the way. We'll begin this exploration with a discussion of how various mutual exclusion primitives like locks are implemented efficiently in modern hardware using benign race conditions. From there, we'll investigate how one can implement non-blocking algorithms and concurrent data structures in a correct and deterministic manner using freely-available open source libraries.
Recomendados
Recomendados
Más contenido relacionado
La actualidad más candente
La actualidad más candente (20)
Similar a Racing To Win: Using Race Conditions to Build Correct and Concurrent Software
Similar a Racing To Win: Using Race Conditions to Build Correct and Concurrent Software (20)
Más de Fastly
Más de Fastly (20)
Último
Último (20)
Racing To Win: Using Race Conditions to Build Correct and Concurrent Software
- 1. Racing To Win Using Race Conditions to Build Correct & Concurrent Software Nathan Taylor | nathan.dijkstracula.net | @dijkstracula
- 2. Racing To Win Using Race Conditions to Build Correct & Concurrent Software Nathan Taylor | nathan.dijkstracula.net | @dijkstracula
- 3. Hi, I’m Nathan. ( @dijkstracula )
- 4. I’m an engineer at
- 6. A problem
- 13. Cache node
- 14. Cache node Process A stack heap text Process B stack heap text Process C stack heap text
- 15. Cache node Process A stack heap text Process B stack heap text Process C stack heap text
- 16. A Persistent, Shared- State Memory Allocator Cache node Process A stack heap text Process B stack heap text Process C stack heap text uSlab
- 17. Slab allocation
- 18. Slab allocation Object Object Object Object Object Object Object Object
- 19. Object Object Object Object Object Object Object Object
- 20. s_alloc(); s_alloc();Object Object Object Object Object Object Object Object s_alloc();
- 21. Object Object Object Object Object Object Object Object
- 22. Object Object Object Object Object Object Object Object
- 23. s_free( ); s_free( ); s_free( ); Object ObjectObject Object Object Object Object Object
- 24. Object Object Object Object Object Object Object Object
- 25. Object Object Object Object Object Object Object Object
- 26. Allocation Protocol • An request to allocate is followed by a response containing an object • A request to free is followed by a response after the supplied object has been released • Allocation requests must not respond with an already- allocated object • A free request must not release an already-unallocated object
- 28. An Execution History void foo() { obj *a = s_alloc(); s_free(a); … }
- 29. An Execution History Time void foo() { obj *a = s_alloc(); s_free(a); … } A(allocate request) B(allocate response) A(free request) B(free response)
- 30. An Execution History Time A(allocate request) B(allocate request) A(allocate response) B(allocate response)
- 31. An Execution History Time A(allocate request) B(allocate request) A(allocate response) B(allocate response) “X happened before Y” => “Y may observe X to have occurred”
- 32. A(allocate response) A(allocate request) B(allocate request) B(allocate response) Time
- 33. A(allocate response) A(allocate request) B(allocate request) B(allocate response) Time
- 34. A(allocate response) A(allocate request) B(allocate request) B(allocate response) A protocol violation! Time
- 35. Time A(allocate response) A(allocate request) B(allocate request) B(allocate response)
- 36. Time A(allocate response) A(allocate request) B(allocate request) B(allocate response)
- 39. A Sequential History Time A(allocate request) A(allocate response) A(free request) A(free response) B(allocate request) B(allocate response)
- 40. A Sequential History Time A(allocate request) A(allocate response) { } A(free request) A(free response) { } B(allocate request) B(allocate response) { }
- 41. A Sequential History Time A(allocate request) A(allocate response) { } A(free request) A(free response) { } B(allocate request) B(allocate response) { }
- 42. A Sequential History Time A(allocate request) A(allocate response) { } A(free request) A(free response) { } B(allocate request) B(allocate response) { }
- 43. obj *allocate(slab *s) { obj *a = s-‐>head; if (a == NULL) return NULL; s-‐>head = a-‐>next; return a; } void free(slab *s, obj *o) { o-‐>next = s-‐>head; s-‐>head = o; }
- 44. obj *allocate(slab *s) { lock(&allocator_lock); obj *a = s-‐>head; if (a == NULL) return NULL; s-‐>head = a-‐>next; unlock(&allocator_lock); return a; } void free(slab *s, obj *o) { lock(&allocator_lock); o-‐>next = s-‐>head; s-‐>head = o; unlock(&allocator_lock); }
- 45. Was the State Locked? Yes Done No Atomic
- 46. Fetch Old Lock State Set State Locked Was old State Locked? Yes Done No Atomic
- 47. Fetch Old Lock State Set State Locked Was old State Locked? Yes Done No Atomic Test And Set Lock
- 48. Test And Set Unlock Set State Unlocked Atomic
- 49. typedef spinlock int; #define LOCKED 1 #define UNLOCKED 0 void lock(spinlock *m) { while (atomic_tas(m, LOCKED) == LOCKED) snooze(); } void unlock(spinlock *m) { atomic_store(m, UNLOCKED); } Many code examples derived from Concurrency Kit http://concurrencykit.org
- 50. void lock(spinlock *m) { while (atomic_tas(m, LOCKED) == LOCKED) snooze(); } A(TAS request) A(TAS response) { }
- 51. A(TAS request) A(TAS response) { } TAS is embedded in Lock
- 52. A(TAS request) A(TAS response) { } A(lock request) A(lock response) Time TAS is embedded in Lock
- 53. A(TAS request) A(TAS response) { } A(lock request) A(lock response) Time TAS & Store can’t be reordered
- 54. A(TAS request) A(TAS response) { } A(lock request) A(lock response) Time B(unlock request) B(unlock response) B(Store request) B(Store response) { } TAS & Store can’t be reordered
- 56. All execution histories All sequentially-consistent execution histories ⊇
- 57. All execution histories All sequentially-consistent execution histories All ???able execution histories ⊇ ⊇
- 58. All execution histories All sequentially-consistent execution histories All linearizable execution histories ⊇ ⊇
- 59. A(TAS request) A(TAS response) { } A(lock request) A(lock response) Time Others can be reordered B(unlock request) B(unlock response) B(Store request) B(Store response) { }
- 60. A(TAS request) A(TAS response) { } A(lock request) A(lock response) Time Others can be reordered B(unlock request) B(unlock response) B(Store request) B(Store response) { }
- 61. void lock(spinlock *m) { while (atomic_tas(m, LOCKED) == LOCKED) snooze(); } void unlock(spinlock *m) { atomic_store(m, UNLOCKED); }
- 63. obj *allocate(slab *s) { lock(&allocator_lock); obj *a = s-‐>head; if (a == NULL) return NULL; s-‐>head = a-‐>next; unlock(&allocator_lock); return a; } void free(slab *s, obj *o) { lock(&allocator_lock); o-‐>next = s-‐>head; s-‐>head = o; unlock(&allocator_lock); }
- 64. void lock(spinlock *m) { while (atomic_tas(m, LOCKED) == LOCKED) snooze(); } void unlock(spinlock *m) { atomic_store(m, UNLOCKED); }
- 65. Spinlock performance millionsoflock acquisitions/sec 15 30 45 60 75 90 Threads 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 87.351 Test and Set
- 66. Spinlock performance millionsoflock acquisitions/sec 15 30 45 60 75 90 Threads 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 Platonic ideal of a spinlock
- 67. Spinlock performance millionsoflock acquisitions/sec 15 30 45 60 75 90 Threads 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 87.351 4.343 Test and Set
- 68. Spinlock performance millionsoflock acquisitions/sec 15 30 45 60 75 90 Threads 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 Test and Set
- 69. Spinlock performance acquisitions/sec 1E+01 1E+02 1E+03 1E+04 1E+05 1E+06 1E+07 1E+08 Threads 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 Test and Set
- 70. typedef spinlock int; #define LOCKED 1 #define UNLOCKED 0 void lock(spinlock *m) { while (atomic_tas(m, LOCKED) == LOCKED) snooze(); }
- 71. typedef spinlock int; #define LOCKED 1 #define UNLOCKED 0 void lock(spinlock *m) { while (atomic_tas(m, LOCKED) == LOCKED) { while (atomic_store(m) == LOCKED) snooze(); } }
- 72. typedef spinlock int; #define LOCKED 1 #define UNLOCKED 0 void lock(spinlock *m) { while (atomic_tas(m, LOCKED) == LOCKED) { while (atomic_store(m) == LOCKED) snooze(); } } Test-and-Test-and-Set
- 73. Lockedalloc/free(10s) 10 100 1,000 10,000 100,000 1,000,000 10,000,000 100,000,000 Threads 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 Test and Set T&T&S
- 74. Spinlock performance Lockedalloc/free(10s) 10 100 1,000 10,000 100,000 1,000,000 10,000,000 100,000,000 Threads 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 Test and Set T&T&S
- 75. typedef spinlock int; #define LOCKED 1 #define UNLOCKED 0 void lock(spinlock *m) { unsigned long backoff, exp = 0; while (atomic_tas(m, LOCKED) == LOCKED) { for (i = 0; i < backoff; i++) snooze(); backoff = (1ULL << exp++); } }
- 76. typedef spinlock int; #define LOCKED 1 #define UNLOCKED 0 void lock(spinlock *m) { unsigned long backoff, exp = 0; while (atomic_tas(m, LOCKED) == LOCKED) { for (i = 0; i < backoff; i++) snooze(); backoff = (1ULL << exp++); } } TAS + backoff
- 77. Lockedalloc/free(10s) 10,000,000 20,000,000 30,000,000 40,000,000 50,000,000 60,000,000 Threads 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 Test and Set T&T&S TAS + EB
- 78. void lock(spinlock *m) { while (atomic_tas(m, LOCKED) == LOCKED) snooze(); } void lock(spinlock *m) { while (atomic_tas(m, LOCKED) == LOCKED) snooze(); } spinlock global_lock = UNLOCKED
- 79. void lock(spinlock *m) { while (atomic_tas(m, LOCKED) == LOCKED) snooze(); } void lock(spinlock *m) { while (atomic_tas(m, LOCKED) == LOCKED) snooze(); } spinlock global_lock = UNLOCKED
- 80. void lock(spinlock *m) { while (atomic_tas(m, LOCKED) == LOCKED) snooze(); } void lock(spinlock *m) { while (atomic_tas(m, LOCKED) == LOCKED) snooze(); } spinlock global_lock = UNLOCKED
- 81. void lock(spinlock *m) { while (atomic_tas(m, LOCKED) == LOCKED) snooze(); } void lock(spinlock *m) { while (atomic_tas(m, LOCKED) == LOCKED) snooze(); } spinlock global_lock = UNLOCKED
- 82. void lock(spinlock *m) { while (atomic_tas(m, LOCKED) == LOCKED) snooze(); } void lock(spinlock *m) { while (atomic_tas(m, LOCKED) == LOCKED) snooze(); } spinlock global_lock = LOCKED
- 83. void lock(spinlock *m) { while (atomic_tas(m, LOCKED) == LOCKED) snooze(); } void lock(spinlock *m) { while (atomic_tas(m, LOCKED) == LOCKED) snooze(); } spinlock global_lock = LOCKED
- 84. void lock(spinlock *m) { while (atomic_tas(m, LOCKED) == LOCKED) snooze(); } void lock(spinlock *m) { while (atomic_tas(m, LOCKED) == LOCKED) snooze(); } spinlock global_lock = LOCKED
- 85. void lock(spinlock *m) { while (atomic_tas(m, LOCKED) == LOCKED) snooze(); } void lock(spinlock *m) { while (atomic_tas(m, LOCKED) == LOCKED) snooze(); } spinlock global_lock = LOCKED
- 86. void lock(spinlock *m) { while (atomic_tas(m, LOCKED) == LOCKED) snooze(); } void lock(spinlock *m) { while (atomic_tas(m, LOCKED) == LOCKED) snooze(); } spinlock global_lock = LOCKED
- 87. void lock(spinlock *m) { while (atomic_tas(m, LOCKED) == LOCKED) snooze(); } void lock(spinlock *m) { while (atomic_tas(m, LOCKED) == LOCKED) snooze(); } spinlock global_lock = LOCKED
- 88. A function is lock-free if at all times at least one thread is guaranteed to be making progress [in the function]. (Herlihy & Shavit)
- 90. // TODO: make this safe and scalable obj *allocate(slab *s) { obj *a = s-‐>head; if (a == NULL) return NULL; s-‐>head = a-‐>next; return a; } // TODO: make this safe and scalable void free(slab *s, obj *o) { o-‐>next = s-‐>head; s-‐>head = o; }
- 92. Compare-And-Swap
- 93. Compare-And-Swap Cmpr and * Old value Destination Address
- 95. Compare-And-Swap Old value New value ≠ Return false = Destination Address Copy to * Return true Cmpr and *
- 96. Compare-And-Swap Old value New value ≠ Return false = Destination Address Return true Atomic Copy to *Cmpr and *
- 97. Atomic i = i+1; void atomic_inc(int *ptr) { int i, i_plus_one; do { i = *ptr; i_plus_one = i + 1; } while (!cas(i, i_plus_one, ptr)); }
- 98. void atomic_inc(int *ptr) { int i, i_plus_one; do { i = *ptr; i_plus_one = i + 1; } while (!cas(i, i_plus_one, ptr)); } Atomic i = i+1;
- 99. void atomic_inc(int *ptr) { int i, i_plus_one; do { i = *ptr; i_plus_one = i + 1; } while (!cas(i, i_plus_one, ptr)); } Atomic i = i+1;
- 100. void atomic_inc(int *ptr) { int i, i_plus_one; do { i = *ptr; i_plus_one = i + 1; } while (!cas(i, i_plus_one, ptr)); } Atomic i = i+1;
- 101. void atomic_inc_mod_32(int *ptr) { int i, new_i; do { i = *ptr; new_i = i + 1; new_i = new_i % 32; } while (!cas(i, new_i, ptr)); } Atomic i = (i+1) % 32;
- 102. TAS using CAS void tas_loop(spinlock *m) { do { ; } while (!cas(UNLOCKED, LOCKED, m)); }
- 103. Read/Modify/Write void atomic_inc_mod_32(int *ptr) { int i, new_i; do { i = *ptr; /* Read */ new_i = fancy_function(); /* Modify */ } while (!cas(i, new_i, ptr)); /* Write */ }
- 104. Read/Modify/Write void atomic_inc_mod_32(int *ptr) { int i, new_i; do { i = *ptr; /* Read */ new_i = fancy_function(); /* Modify */ } while (!cas(i, new_i, ptr)); /* Write */ /* (or retry) */ }
- 105. obj *allocate(slab *s) { obj *a, *b; do { a = s-‐>head; if (a == NULL) return NULL; b = a-‐>next; } while (!cas(a, b, &s-‐>head )); return a; } slab head A B …
- 106. obj *allocate(slab *s) { obj *a, *b; do { a = s-‐>head; if (a == NULL) return NULL; b = a-‐>next; } while (!cas(a, b, &s-‐>head )); return a; } slab head A B …
- 107. obj *allocate(slab *s) { obj *a, *b; do { a = s-‐>head; if (a == NULL) return NULL; b = a-‐>next; } while (!cas(a, b, &s-‐>head )); return a; } A B …slab head
- 108. B … slab head A obj *allocate(slab *s) { obj *a, *b; do { a = s-‐>head; if (a == NULL) return NULL; b = a-‐>next; } while (!cas(a, b, &s-‐>head )); return a; }
- 109. obj *allocate(slab *s) { obj *a, *b; do { a = s-‐>head; if (a == NULL) return NULL; b = a-‐>next; } while (!cas( , , )); return a; } slab head a a b Cmpr and * &s->head A B … b a
- 110. slab head Cmpr and Z obj *allocate(slab *s) { obj *a, *b; do { a = s-‐>head; if (a == NULL) return NULL; b = a-‐>next; } while (!cas( , , )); return a; } a a b &s->head b a
- 111. slab head Z A B Cmpr and obj *allocate(slab *s) { obj *a, *b; do { a = s-‐>head; if (a == NULL) return NULL; b = a-‐>next; } while (!cas( , , )); return a; } a a b &s->head b a
- 112. slab head B … Cmpr and obj *allocate(slab *s) { obj *a, *b; do { a = s-‐>head; if (a == NULL) return NULL; b = a-‐>next; } while (!cas( , , )); return a; } a a b &s->head b a
- 113. void free(slab *s, obj *o) { do { obj *t = s-‐>head; o-‐>next = t; } while (!cas(t, o, &s-‐>head)); } B …slab head
- 114. void free(slab *s, obj *o) { do { obj *t = s-‐>head; o-‐>next = t; } while (!cas(t, o, &s-‐>head)); } slab head A B …
- 115. A B Cslab head
- 116. A B Cslab head
- 117. obj *allocate(slab *s) { obj *a, *b; do { a = s-‐>head; if (a == NULL) return NULL; b = a-‐>next; } while (!cas(a, b, &s-‐>head)); return a; } A B Cslab head
- 118. obj *allocate(slab *s) { obj *a, *b; do { a = s-‐>head; if (a == NULL) return NULL; b = a-‐>next; } while (!cas(a, b, &s-‐>head)); return a; } A B Cslab head
- 119. A B C obj *allocate(slab *s) { obj *a, *b; do { a = s-‐>head; if (a == NULL) return NULL; b = a-‐>next; } while (!cas(a, b, &s-‐>head)); return a; } slab head
- 120. A B C obj *allocate(slab *s) { obj *a, *b; do { a = s-‐>head; if (a == NULL) return NULL; b = a-‐>next; } while (!cas(a, b, &s-‐>head)); return a; } slab head
- 121. A B C some_object = allocate(&shared_slab); slab head obj *allocate(slab *s) { obj *a, *b; do { a = s-‐>head; if (a == NULL) return NULL; b = a-‐>next; } while (!cas(a, b, &s-‐>head)); return a; }
- 122. A B C some_object = allocate(&shared_slab); slab head obj *allocate(slab *s) { obj *a, *b; do { a = s-‐>head; if (a == NULL) return NULL; b = a-‐>next; } while (!cas(a, b, &s-‐>head)); return a; }
- 123. B C A slab head some_object = allocate(&shared_slab); obj *allocate(slab *s) { obj *a, *b; do { a = s-‐>head; if (a == NULL) return NULL; b = a-‐>next; } while (!cas(a, b, &s-‐>head)); return a; }
- 124. B C A slab head some_object = allocate(&shared_slab); obj *allocate(slab *s) { obj *a, *b; do { a = s-‐>head; if (a == NULL) return NULL; b = a-‐>next; } while (!cas(a, b, &s-‐>head)); return a; }
- 125. B C another_obj = allocate(&shared_slab); A slab head some_object = allocate(&shared_slab); obj *allocate(slab *s) { obj *a, *b; do { a = s-‐>head; if (a == NULL) return NULL; b = a-‐>next; } while (!cas(a, b, &s-‐>head)); return a; }
- 126. B C another_obj = allocate(&shared_slab); A slab head some_object = allocate(&shared_slab); obj *allocate(slab *s) { obj *a, *b; do { a = s-‐>head; if (a == NULL) return NULL; b = a-‐>next; } while (!cas(a, b, &s-‐>head)); return a; }
- 127. C A B slab head another_obj = allocate(&shared_slab); some_object = allocate(&shared_slab); obj *allocate(slab *s) { obj *a, *b; do { a = s-‐>head; if (a == NULL) return NULL; b = a-‐>next; } while (!cas(a, b, &s-‐>head)); return a; }
- 128. C A B slab head another_obj = allocate(&shared_slab); some_object = allocate(&shared_slab); obj *allocate(slab *s) { obj *a, *b; do { a = s-‐>head; if (a == NULL) return NULL; b = a-‐>next; } while (!cas(a, b, &s-‐>head)); return a; }
- 129. B C A slab head another_obj = allocate(&shared_slab); free(&shared_slab, some_object); obj *allocate(slab *s) { obj *a, *b; do { a = s-‐>head; if (a == NULL) return NULL; b = a-‐>next; } while (!cas(a, b, &s-‐>head)); return a; }
- 130. B C A slab head another_obj = allocate(&shared_slab); free(&shared_slab, some_object); obj *allocate(slab *s) { obj *a, *b; do { a = s-‐>head; if (a == NULL) return NULL; b = a-‐>next; } while (!cas(a, b, &s-‐>head)); return a; }
- 131. B A Cslab head another_obj = allocate(&shared_slab); obj *allocate(slab *s) { obj *a, *b; do { a = s-‐>head; if (a == NULL) return NULL; b = a-‐>next; } while (!cas(a, b, &s-‐>head)); return a; } free(&shared_slab, some_object);
- 132. B A Cslab head another_obj = allocate(&shared_slab); obj *allocate(slab *s) { obj *a, *b; do { a = s-‐>head; if (a == NULL) return NULL; b = a-‐>next; } while (!cas(a, b, &s-‐>head)); return a; } free(&shared_slab, some_object);
- 133. B A Cslab head another_obj = allocate(&shared_slab); obj *allocate(slab *s) { obj *a, *b; do { a = s-‐>head; if (a == NULL) return NULL; b = a-‐>next; } while (!cas(a, b, &s-‐>head)); return a; } free(&shared_slab, some_object);
- 134. B A Cslab head another_obj = allocate(&shared_slab); obj *allocate(slab *s) { obj *a, *b; do { a = s-‐>head; if (a == NULL) return NULL; b = a-‐>next; } while (!cas(a, b, &s-‐>head)); return a; } free(&shared_slab, some_object);
- 135. free(&shared_slab, some_object); B B Cslab head A another_obj = allocate(&shared_slab); obj *allocate(slab *s) { obj *a, *b; do { a = s-‐>head; if (a == NULL) return NULL; b = a-‐>next; } while (!cas(a, b, &s-‐>head)); return a; }
- 136. free(&shared_slab, some_object); B B Cslab head A another_obj = allocate(&shared_slab); obj *allocate(slab *s) { obj *a, *b; do { a = s-‐>head; if (a == NULL) return NULL; b = a-‐>next; } while (!cas(a, b, &s-‐>head)); return a; }
- 137. free(&shared_slab, some_object); B B Cslab head A another_obj = allocate(&shared_slab); obj *allocate(slab *s) { obj *a, *b; do { a = s-‐>head; if (a == NULL) return NULL; b = a-‐>next; } while (!cas(a, b, &s-‐>head)); return a; }
- 138. The ABA Problem “A reference about to be modified by a CAS changes from a to b and back to a again. As a result, the CAS succeeds even though its effect on the data structure has changed and no longer has the desired effect.” —Herlihy & Shavit, p. 235
- 139. obj *allocate(slab *s) { obj *a, *b; do { a = s-‐>head; if (a == NULL) return NULL; b = a-‐>next; } while (!cas(a, b, &s-‐>head)); return a; } A B …slab head 166
- 140. obj *allocate(slab *s) { slab orig, update; do { orig.gen = s.gen; orig.head = s.head; if (!orig.head) return NULL; update.gen = orig.gen + 1; update.head = orig.head-‐>next; } while (!dcas(&orig, &update, s)); return orig.head; } A B …slab head 166
- 141. free(slab *s, obj *o) { do { obj *t = s-‐>head; o-‐>next = t; } while (!cas(t, o, &s-‐>head)); }
- 142. obj *allocate(slab *s) { lock(&allocator_lock); obj *a = s-‐>head; if (a == NULL) return NULL; s-‐>head = a-‐>next; unlock(&allocator_lock); return a; } void free(slab *s, obj *o) { lock(&allocator_lock); o-‐>next = s-‐>head; s-‐>head = o; unlock(&allocator_lock); }
- 143. slab head A B … obj *o = allocate(&shared_slab); obj *o = allocate(&shared_slab);
- 144. slab head B … obj *o = allocate(&shared_slab); obj *o = allocate(&shared_slab); A A
- 145. slab head B … obj *o = allocate(&shared_slab); obj *o = allocate(&shared_slab); A A Memory barriers
- 146. lock(&allocator_lock); obj *a = s-‐>head; … lock(&allocator_lock); obj *a = s-‐>head; …
- 147. while (atomic_tas(m, LOCKED) == LOCKED) snooze(); obj *a = s-‐>head; … while (atomic_tas(m, LOCKED) == LOCKED) snooze(); obj *a = s-‐>head; …
- 148. LDREX R5, [m] ; TAS: fetch. . . STREXEQ R5, LOCKED, [m] ; TAS: . . . and set CMPEQ R5, #0 ; Did we succeed? LDR R0, [R1, 4] ; a = s-‐>head BEQ lock_done ; Yes: we are all done BL snooze ; No: Call snooze()… B lock_loop ; …then loop again lock_done: B LR ; return ;;;; IN lock() lock_loop: ;;;; IN allocate()
- 149. LDREX R5, [m] ; TAS: fetch. . . STREXEQ R5, LOCKED, [m] ; TAS: . . . and set CMPEQ R5, #0 ; Did we succeed? LDR R0, [R1, 4] ; a = s-‐>head BEQ lock_done ; Yes: we are all done BL snooze ; No: Call snooze()… B lock_loop ; …then loop again lock_done: B LR ; return ;;;; IN allocate() ;;;; IN lock() lock_loop:
- 150. LDR R0, [R1, 4] ; a = s-‐>head BEQ lock_done ; Yes: we are all done BL snooze ; No: Call snooze()… B lock_loop ; …then loop again lock_done: B LR ; return ;;;; IN allocate() LDREX R5, [m] ; TAS: fetch. . . STREXEQ R5, LOCKED, [m] ; TAS: . . . and set CMPEQ R5, #0 ; Did we succeed? ;;;; IN lock() lock_loop:
- 151. McKenney , p. 504
- 152. McKenney , p. 504
- 153. McKenney , p. 504
- 154. McKenney , p. 504
- 156. LDREX R5, [m] ; TAS: fetch. . . STREXEQ R5, LOCKED, [m] ; TAS: . . . and set CMPEQ R5, #0 ; Did we succeed? LDR R0, [R1, 4] ; a = s-‐>head BEQ lock_done ; Yes: we are all done BL snooze ; No: Call snooze()… B lock_loop ; …then loop again lock_done: B LR ; return ;;;; IN allocate() ;;;; IN lock() lock_loop:
- 157. LDREX R5, [m] ; TAS: fetch. . . STREXEQ R5, LOCKED, [m] ; TAS: . . . and set CMPEQ R5, #0 ; Did we succeed? LDR R0, [R1, 4] ; a = s-‐>head BEQ lock_done ; Yes: we are all done BL snooze ; No: Call snooze()… B lock_loop ; …then loop again lock_done: B LR ; return ;;;; IN allocate() ;;;; IN lock() lock_loop:
- 159. obj *a = s-‐>head; lock(&allocator_lock); … obj *a = s-‐>head; lock(&allocator_lock); …
- 160. lock(&allocator_lock); obj *a = s-‐>head; … lock(&allocator_lock); obj *a = s-‐>head; …
- 161. lock(&allocator_lock); < -‐ -‐ -‐ -‐ -‐ -‐ -‐ -‐ -‐ -‐> obj *a = s-‐>head; … lock(&allocator_lock); < -‐ -‐ -‐ -‐ -‐ -‐ -‐ -‐ -‐ -‐> obj *a = s-‐>head; …
- 163. LDREX R5, [m] ; TAS: fetch. . . STREXEQ R5, LOCKED, [m] ; TAS: . . . and set CMPEQ R5, #0 ; Did we succeed? BEQ lock_done ; Yes: we are all done BL snooze ; No: Call snooze()… B lock_loop ; …then loop again lock_done: DMB ; Ensure all previous reads ; have been completed B LR ; return ;;;; IN unlock() MOV R0, UNLOCKED DMB ; Ensure all previous reads have ; been completed STR R0, LR ;;;; IN lock() lock_loop:
- 164. nathan~$ cat /proc/cpuinfo | grep "physical.*0" | wc -‐l 16 nathan~$ cat /proc/cpuinfo | grep "model name" | uniq model name : Intel(R) Xeon(R) CPU E5-‐2690 0 @ 2.90GHz Allocator performance
- 165. MillionsofAlloc/free pairs/sec 10 20 30 40 50 60 Threads 1 20.56822.392 50.52951.23452.721 T&S T&S-EB T&T&S CAS pthread_mutex Allocator Throughput
- 166. MillionsofAlloc/free pairs/sec 10 20 30 40 50 60 Threads 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 TAS T&T&S TAS + EB Concurrent Allocator pthread Allocator Throughput
- 167. Allocator latency
- 173. The lyf so short, the CAS so longe to lerne • Cache coherency and NUMA architecture • Transactional memory
- 174. #thoughtleadership
- 175. a safe race? When is a race
- 177. “lock-free programming is hard; let’s go ride bikes”?
- 178. “lock-free programming is hard; let’s go ride bikes”? • high-level performance necessitates an understanding of low level performance
- 179. “lock-free programming is hard; let’s go ride bikes”? • high-level performance necessitates an understanding of low level performance • your computer is a distributed system
- 180. “lock-free programming is hard; let’s go ride bikes”? • high-level performance necessitates an understanding of low level performance • your computer is a distributed system • (optional third answer: it’s real neato)
- 187. Come see us at the booth! Nathan Taylor | nathan.dijkstracula.net | @dijkstracula Thanks credits, code, and additional material at https://github.com/dijkstracula/Surge2015/