Optimizing Swift Collections

In this talk, I describe several ways to implement the same simple ordered set protocol in Swift, demonstrating how the language supports a number of surprisingly different approaches to programming. At every step, we trade extra complexity for improved runtime performance, ending on an implementation that is ludicrously fast but also quite difficult to handle.

Presented on dotSwift 2017 in January 2017.

The talk has a repository on GitHub, with full source code and further details:

https://github.com/lorentey/OptimizingCollections

Transcript

1. Optimizing Swift Collections Károly Lőrentey @lorentey
6. NSOrderedSet
10. 1 4 8 9 11 12 14 16 17
11. 1 4 8 9 11 12 14 16 17 5
12. 1 4 5 8 9 11 12 14 16 17
18. 40x
19. 40x
21. 9 1 8 11 16 5 12 4 14 17 left right
22. x a b y c d Empty Tree Black node Red node
23. .empty .node(color, a, x, b) x a b y c d
24. 9 1 8 11 16 5 12 4 14 17 4 2 3 5 1 8 9 a b 6 7
25. 9 1 8 11 16 5 12 4 14 17 4 2 3 5 1 8 9 a b 6 7
26. 9 1 8 11 16 5 12 4 14 17
27. 9 1 8 11 16 5 12 4 14 17
28. 9 1 8 11 16 5 12 4 14 17
29. 9 1 8 11 16 5 12 4 14 17
38. z x y a b c d x z y a b c d z y x a b c d x y z a b c d
39. x z y a b c d z x y a b c d x z y a b c d z y x a b c d x y z a b c d
45. 2.5x
46. 2.5x
47. 2.5x ???
48. 9 11 12 5 8 1 4 16 14 17 t t.insert(10)
49. 9 11 12 5 8 1 4 16 14 17 t 10 t.insert(10)
50. 9 11 12 5 8 1 4 16 14 17 t 10 t.insert(10)
51. 9 11 12 5 8 1 4 16 14 17 t 10 t.insert(10)
52. 9 11 12 5 8 1 4 16 14 17 t 10 t.insert(10)
53. 9 11 12 5 8 1 4 16 14 17 t 10 11ʹ t.insert(10)
54. 9 11 12 5 8 1 4 16 14 17 t 10 11ʹ 12ʹ t.insert(10)
55. 9 11 12 5 8 1 4 16 14 17 t 10 11ʹ t.inserting(10) 9ʹ 12ʹ t.insert(10)
56. 9 12 11 5 8 1 4 16 14 17 t 10 11ʹ 12ʹ t.inserting(10) 9ʹ t.insert(10)
57. 9 12 11 5 8 1 4 16 14 17 t 10 11ʹ 12ʹ 9ʹ t.insert(10)
58. 12 11 5 8 1 4 16 14 17 t 10 11ʹ 12ʹ 9ʹ t.insert(10)
59. 11 5 8 1 4 16 14 17 t 10 11ʹ 12ʹ 9ʹ t.insert(10)
60. 5 8 1 4 16 14 17 t 10 11ʹ 12ʹ 9ʹ t.insert(10)
61. 5 8 1 4 16 14 17 t 10 11ʹ 12ʹ 9ʹ t.insert(10)
62. 5 8 1 4 16 14 17 t 10 11ʹ 12ʹ 9ʹ 3 allocations 3 deallocations 10 atomic refCount ops t.insert(10)
63. 9 11 12 5 8 1 4 16 14 17 t.insert(10) t
64. 9 11 12 5 8 1 4 16 14 17 t.insert(10) 10 t
65. 9 11 12 5 8 1 4 16 14 17 t.insert(10) 10 t 10
66. 9 11 12 5 8 1 4 16 14 17 t.insert(10) 10 1 allocation No deallocations No atomic refCount ops t 10
67. 9 11 12 5 8 1 4 16 14 17 t.insert(10) 10 1 allocation No deallocations No atomic refCount ops t 10
68. COW
69. opy C Write on
71. COWTree root: COWNode color: value: 10 left: right: COWNode COWNode
80. !
83. 4.5x
84. 4.5x ???
87. O(n) Linear Time
88. O(1) Constant Time O(n) Linear Time
89. O(1) Constant Time O(n) Linear Time 4096
91. Ints 4096
92. 4096 × 8 bytes
93. 4096 × 8 bytes = 32 kB
94. 4096 × 8 bytes = 32 kB = size of L1 cache
95. 1 2 3 4
96. 1 2 3 4 5
97. 1 2 3 4 5 4
98. 1 2 3 4 5 4
99. 1 2 3 4 5 4 7
100. 1 2 3 4 5 4 7 8 9
101. 1 2 3 4 5 4 7 8 9
102. 1 2 3 4 5 4 7 8 9
103. 1 2 3 4 5 4 7 8 9 10
104. 1 2 3 4 5 4 7 8 9 10 11 12
105. 1 2 3 4 b 4 7 8 9 10 11 12 5 a c
106. 1 2 3 4 b 4 7 8 9 10 11 12 5 a c
107. 1 2 3 4 b 4 7 8 9 10 11 12 5 a c
108. 1 2 3 4 b 4 7 8 9 10 11 12 5 a c
109. 1 2 3 4 b 4 7 8 9 10 11 12 5 a c
110. 9 c 1 2 3 4 b 4 7 8 1 1 1 5a
111. 9 c d e f g h i 1 2 3 4 b 4 7 8 1 1 1 5a 34 7 8 1 2 b 4 1 2 b 4 1 3 47 8 1 2 b 4 1 2 b 4 1 2 Z 4 1 j
112. 9 c d e f g h i 1 2 3 4 b 4 7 8 1 1 1 5a 34 7 8 1 2 b 4 1 2 b 4 1 3 47 8 1 2 b 4 1 2 b 4 1 2 Z 4 1 j
113. 9 c d e f g h i 1 2 3 4 b 4 7 8 1 1 1 5a 34 7 8 1 2 b 4 1 2 b 4 1 3 47 8 1 2 b 4 1 2 b 4 1 2 Z 4 1 j B-tree
123. 3.5x 4.5x 2.5x
125. 1 2 34 54 7 8 9 1 2 34 54 7 8 9 Reuse an existing Collection ♻
126. 1 2 34 54 7 8 9 1 2 34 54 7 8 9 Reuse an existing Collection ♻ Functional programming 
 with algebraic data types λ
127. 1 2 34 54 7 8 9 1 2 34 54 7 8 9 Reuse an existing Collection ♻ Imperative rewrite with
 COW optimization # Functional programming 
 with algebraic data types λ
128. 9 c de f gh i j 1 2 34 54 7 8 9 1 2 34 54 7 8 9 Reuse an existing Collection ♻ Cache-aware data structures and unsafe operations ⚙ Imperative rewrite with
 COW optimization # Functional programming 
 with algebraic data types λ
129. Reuse an existing Collection ♻
130. import BTree
131. import BTree
132. 22m 40s SortedArray NSOrderedSet AlgebraicTree COWTree BTree/1024 Time spent inserting 3 million random values
133. 22m 40s SortedArray NSOrderedSet AlgebraicTree COWTree BTree/1024 1.8s 7s 22s 14m 54s 22m 36s Time spent inserting 3 million random values
135. Thank You!
136. Thank You! lorentey
137. Thank You! lorentey @lorentey
139. Acknowledgements Cow picture courtesy of freegreatpicture.com Puppy pictures courtesy of Floppy the Beagle