Server Memory - BuildStuff Ukraine 2019

Transcript

1. None

2. @listochkin + fb.com/tektonna

3. Memory Management

4. A bit deeper than most talks on similar subject

5. None

6. CPU Caches RAM Swap Disk Network

7. Operating System

8. None

9. Pages 4k

10. Virtual

11. 3 processes [0 … ∞] [0 … ∞] [0 …

    ∞]

12. Allocate page But not really

13. None

14. malloc calloc realloc

15. None

16. pointer = malloc(size)

17. free(pointer)

18. malloc malloc malloc free malloc free free

19. Forget to free Free the wrong pointer Read from unallocated

    memory

20. Can we do better?

21. Reference Counting

22. None

23. const players = [ … ] players.forEach(p => makeMove(p))

24. a = … doX(a) yield a

25. if ( … ) { let a = … …

    }

26. C++: std::shared_ptr

27. None

28. Optimizations

29. Delayed Counting

30. Don’t count local references

31. function (user) { let email = user.email; … … }

32. Delayed deallocation

33. Reference Cycles

34. None
35. None

36. Compile time ref-counting

37. Several types of pointers

38. Ownership

39. Forbids cycles by forbidding several owning references at the same

    time
40. None
41. None
42. None

43. Perl 1987 - today Python 1991 - 2001 PHP 1994

    - 2009

44. Multiprocess deploys

45. fork

46. Copy on Write

47. Worker Accepts requests As memory use raises: Stop accepting requests

    Complete in-flight requests Terminate

48. Master Keep track on workers Start new workers Signal them

    to terminate when memory pressure is high

49. “Pre-fork”

50. Load the framework Load app code Run full GC ----

    Start forking process to accept requests
51. None
52. None

53. “Processes” Supervision Trees

54. Tracing GC

55. 1959 60 years ago

56. Start at Root references Follow all references Build a live

    objects tree Delete all objects not part of the tree

57. Ruby Java JavaScript Lua Go Default Choice

58. Roots?

59. Constants Global variables Local variables Closures Thread-locals ...

60. Mark & Sweep vs Mark-Compact

61. None

62. Team Sweep: Go Ruby* Lua Embedded JS engines Erlang

63. Pros: Pointers don’t change Native extensions Easier to implement

64. Team Compact: Java JavaScript Ruby* Haskell

65. Pros: Less memory fragmentation over time Cons: Harder Takes longer

    to do a GC

66. Incremental Marking

67. 3-colored algorithm by Dijkstra™

68. None
69. None
70. None
71. None
72. None

73. What is a barrier? if statement

74. CPU branch predictor

75. Parallel marking

76. None

77. Lazy Sweep

78. Generational

79. Temporary data data.filter( … ).map( … ) log(`… ${data}`) JSON.stringify(result)

80. Major vs Minor GC

81. Pointers from Old objects to new objects

82. Remembered Set

83. Major GC Scan roots only Bigger object graph

84. Minor GC Scan roots + RS Small object graph

85. None

86. Modern GCs are hybrid

87. None
88. None

89. C-extensions

90. Can’t move objects if their references are passed to an

    extension

91. Can’t add WB

92. RGen GC

93. 2 types of objects WB-protected WB-unprotected

94. WBu are never OldGen

95. OldGen -> WBu WBu to Remembered Set

96. Mark WBus on every minor GC

97. Adding compaction for WBp

98. None

99. V8

100. Minor GC Parallel

101. Major Parallel marking Parallel / Concurrent Compact || Sweep

102. None

103. When to trigger GC?

104. Out-of-Bounds GC

105. Firefox Run GC in background tabs first instead of current

     tab

106. Chrome Animation frame

107. OS Pages pre-forked processes malloc zones GC Regions Remembered Sets

     Barriers
108. None
109. None

110. Practical Advice

111. Memory Leaks Multiprocess Deploys

112. Throughput ⇔ Latency

113. Throughput Batch jobs Queues

114. Latency Requests Messaging

115. Keep different types of workloads in different processes

116. More memory Swap with SSDs

117. Experiment and measure

118. @listochkin Hire Me!