1. Flavio W. Brasil
@flaviowbrasil
Twitter, VM Team
Scala ❤ Graal

2. #TwitterVMTeam

3. Scala’s
performance
challenges
Why Twitter
adopted Graal
New
optimizations
1 #Deploy 2 #Profile 3 #Optimize
Agenda

4. #Deploy
Why Twitter adopted Graal

5. What is Graal?

6. Graal != GraalVM

7. Graal != GraalVM

9. What is a JIT compiler?

10. Source code

11. Source code Bytecode

12. Source code Bytecode Native

14. Performance
Ease of change
Why Graal

17. C2 is a 20yo
codebase in C++

18. Graal is much
easier to change

19. #Profile
Scala’s performance challenges

20. Scala is not Java
(but Java is becoming Scala)

22. Composable APIs
Scala’s performance nightmare

23. Composable APIs are
interpreted languages

39. Why does a single
indirection have
this effect?

44. 👍

45. 😕

46. Interpreted languages are
difficult to optimize

47. Embedded
Interpreted languages are
even more
difficult to optimize

48. #Optimize
New optimizations by the #TwitterVMTeam

49. First target: Future

59. Synchronization

70. Opaque

71. CAS Virtualization
https://github.com/oracle/graal/pull/636

76. Q4/2018
Results

77. * Includes code changes
https://github.com/twitter/util/commit/3245a8
-16% alloc/s-5% CPU
CAS Optimizations
(regular load)

78. All optimizations
(stress test)

79. All optimizations
(stress test)

80. All optimizations
(stress test)

81. Second target:
itable stub calls

82. What is an itable stub?

83. Virtual call

84. Virtual call

85. Virtual call

86. Virtual call
Object
getClass 1
toString 2

87. Virtual call
Object
getClass 1
toString 2
Mammal
getClass 1
toString 2

88. Virtual call
Object
getClass 1
toString 2
Mammal
getClass 1
toString 2
speak 3

89. Virtual call
Object
getClass 1
toString 2
Mammal
getClass 1
toString 2
speak 3
Human
getClass 1
toString 2
speak 3

90. Virtual call
Object
getClass 1
toString 2
Mammal
getClass 1
toString 2
speak 3
Human
getClass 1
toString 2
speak 3
speak(String) 4

91. Interface call

92. Interface call

93. Interface call

94. Interface call
Human
Named 1
Mammal 2

95. Interface call
Human
Named 1
Mammal 2
Human/
Mammal
speak 2
Human/
Named
name 1

96. Interface call
Human
Named 1
Mammal 2
Human/
Mammal
speak 2
Human/
Named
name 1
❓

97. Interface call
Human
Named 1
Mammal 2
Human/
Mammal
speak 2
Human/
Named
name 1
❌

98. Interface call
Human
Named 1
Mammal 2
Human/
Mammal
speak 2
Human/
Named
name 1
❌
❓

99. Interface call
Human
Named 1
Mammal 2
Human/
Mammal
speak 2
Human/
Named
name 1
❌
✅

100. Interface call
Human
Named 1
Mammal 2
Human/
Mammal
speak 2
Human/
Named
name 1
❌
✅
✅

101. How can we optimize it?

103. Data-driven approach

105. Top invokes

106. Common offsets

107. Common offsets

108. Top invokes Common offsets

109. What if a new class is
observed?

110. Call
A
B

111. Call
A
B
C

112. Deoptimization
is challenging

113. Single method

114. Common super class

115. Fingerprint

118. Q1/2019
Initial results

119. Invoke interface optimization
(stress test)
Max RPS Max QPS
Baseline 735 3762
Optimized 1349 4548
Improvement 83% 20%

120. Before After
Invoke interface optimization
(stress test)

121. Before
After

122. Performance ✅
Ease of change ✅
Why Graal

123. Thank you!