[My personal, biased view of] The Last Five Years of Energy Consumption Research

Transcript

1. The Last Five Years of Energy Consumption Research Gustavo Pinto My personal, biased view of @gustavopinto gpinto@ufpa.br
2. 2012..
7. What I can do about it?
8. 2013 Ambitious Plan “propose a catalog of refactorings targeting some languages of the JVM platform.” SPLASH DOC SYMP’13
9. 2013 Ambitious Plan “propose a catalog of refactorings targeting some languages of the JVM platform.” SPLASH DOC SYMP’13
11. Refactorings Multi-core Systems Catalogue 2 JVM PL
12. 2013 2M+ Users 5M+ Questions 10M+ Answers 50GB+ of data MSR’14 2014
13. 13 5M Questions Manual Filter Final Data Automatic Filter
14. 14 5M Questions Manual Filter Final Data Automatic Filter 615 Questions 1,197 Answers
15. 15 5M Questions Automatic Filter Manual Filter Final Data from 2008 to 2013 325 Questions 558 Answers Base Group
16. Research Questions 16 • RQ1: What are the most common energy-related problems faced by software developers? • RQ2: What are the main causes for software energy consumption problems? • RQ3: What solutions do developers employ or recommend to save energy?
17. Energy-Related Problems 17 • Measurements (59/97 — Q/A) • General Knowledge (40/84 — Q/A) • Code design (36/133 — Q/A) • Context-specific (83/110 — Q/A) • Noise (107/134 — Q/A)
18. 18 “I want to measure the energy consumption of my own application (which I can modify) [...] on Windows CE 5.0 and Windows Mobile 5/6. Is there some kind of API for this?” • Measurements (59/97 — Q/A) • General Knowledge (40/84 — Q/A) • Code design (36/133 — Q/A) • Context-specific (83/110 — Q/A) • Noise (107/134 — Q/A)
19. 19 “Can a code optimized for least MCPS be guaranteed to have least power consumption as well?” • Measurements (59/97 — Q/A) • General Knowledge (40/84 — Q/A) • Code design (36/133 — Q/A) • Context-specific (83/110 — Q/A) • Noise (107/134 — Q/A)
20. 20 “Are there any s/w high level design considerations [...] to make the code as power efficient as possible?” • Measurements (59/97 — Q/A) • General Knowledge (40/84 — Q/A) • Code design (36/133 — Q/A) • Context-specific (83/110 — Q/A) • Noise (107/134 — Q/A)
21. 21 • Measurements (59/97 — Q/A) • General Knowledge (40/84 — Q/A) • Code design (36/133 — Q/A) • Context-specific (83/110 — Q/A) • Noise (107/134 — Q/A) — Highest popularity — Highest A per Q ratio — Highest success rate Energy-Related Problems
22. Energy-Related Causes 22 • Unnecessary resource usage (49 occurrences) • Fault GPS behavior (42 occurrences) • Background activities (40 occurrences) • Excessive synchronization (32 occurrences) • Background wallpapers (17 occurrences) • Advertisement (11 occurrences)
23. 23 • Unnecessary resource usage (49 occurrences) • Fault GPS behavior (42 occurrences) • Background activities (40 occurrences) • Excessive synchronization (32 occurrences) • Background wallpapers (17 occurrences) • Advertisement (11 occurrences) “to have a background application that monitors device usage, identifies unused/idle resources, and acts appropriately”
24. 24 • Unnecessary resource usage (49 occurrences) • Fault GPS behavior (42 occurrences) • Background activities (40 occurrences) • Excessive synchronization (32 occurrences) • Background wallpapers (17 occurrences) • Advertisement (11 occurrences) “When there are bugs that keep the GPS turned on too long they go to the top of the list to get fixed”
25. Energy-Related Solutions 25 • Keep IO to a minimum (29 occurrences) • Bulk operations (24 occurrences) • Avoid polling (17 occurrences) • Hardware Coordination (11 occurrences) • Concurrent Programming (9 occurrences) • Race to idle (7 occurrences)
26. 26 • Keep IO to a minimum (29 occurrences) • Bulk operations (24 occurrences) • Avoid polling (17 occurrences) • Hardware Coordination (11 occurrences) • Concurrent Programming (9 occurrences) • Race to idle (7 occurrences) “do not flood the output stream with null values”
27. 27 • Keep IO to a minimum (29 occurrences) • Bulk operations (24 occurrences) • Avoid polling (17 occurrences) • Hardware Coordination (11 occurrences) • Concurrent Programming (9 occurrences) • Race to idle (7 occurrences) “Don’t transfer say 1 file, and then wait for a bit to do another transfer. Instead, transfer right after the other.”
28. Do researchers agree? 28 • Keep IO to a minimum (29 occurrences) • Bulk operations (24 occurrences) • Avoid polling (17 occurrences) • Hardware Coordination (11 occurrences) • Concurrent Programming (9 occurrences) • Race to idle (7 occurrences)
29. Do researchers agree? 29 • Keep IO to a minimum (29 occurrences) • Bulk operations (24 occurrences) • Avoid polling (17 occurrences) • Hardware Coordination (11 occurrences) • Concurrent Programming (9 occurrences) • Race to idle (7 occurrences)
30. 2014 30 • Explicit threading (the Thread-style): Using the java.lang.Thread class • Thread pooling (the Executor-style): Using the java.util.concurrent.Executor framework • Working Stealing (the ForkJoin-style): Using the java.util.concurrent.ForkJoin framework OOPSLA’14
31. 31 • Embarrassingly parallel: spectralnorm, sunflow, n-queens • Leaning parallel: xalan, knucleotide, tomcat • Leaning serial: mandelbrot, largestImage • Embarrassingly serial: h2 Benchmarks
32. 32 • Embarrassingly parallel: spectralnorm, sunflow, n-queens • Leaning parallel: xalan, knucleotide, tomcat • Leaning serial: mandelbrot, largestImage • Embarrassingly serial: h2 Benchmarks Micro-benchmarks DaCapo benchmarks
33. 33 Experimental Environment A 2×16-core AMD CPUs, running Debian Linux, 64GB of memory, JDK version 1.7.0 11, build 21, “ondemand” governor
34. 34 Experimental Environment A 2×16-core AMD CPUs, running Debian Linux, 64GB of memory, JDK version 1.7.0 11, build 21, “ondemand” governor
35. 35 Experimental Environment A 2×16-core AMD CPUs, running Debian Linux, 64GB of memory, JDK version 1.7.0 11, build 21.
36. 36 Experimental Environment A 2×16-core AMD CPUs, running Debian Linux, 64GB of DDR3 1600 memory, and JDK version 1.7.0 11, build 21.
37. Methodology 37 Starting from the Thread-style
38. Methodology 38 Then to the Executors-style
39. Methodology 39 And finally, the ForkJoin-style
40. Energy Consumption When Varying the Number of Threads 40
41. 41 The Λ Curve
42. 42 The Λ Curve
43. 43 The Λ Curve
44. 44 More cores idle CPU frequency at a lower level The Λ Curve
45. 45 More cores idle CPU frequency at a lower level More threads used, program completes sooner The greater the ratio between speedup and power, the steeper the \ The Λ Curve
46. Copying vs Sharing 46
47. Copying vs Sharing 47 Copying
48. Copying vs Sharing 48 Copying Sharing
49. Copying vs Sharing 49 Copying Sharing ±15% of energy savings!
50. Copying vs Sharing 50 Copying Sharing ±15% of energy savings! Interesting. Thanks! Doug Lea
52. 2015 52 4Mi+ Users 19Mi+ Repositories MSR’15
53. 53 19M Repos
54. 54 19M Repos Query on
55. 55 19M Repos Query on
56. 56 19M Repos 2,189 Commits Query on
57. 57 19M Repos Automatic Filter 1,005 Commits Query on
58. 58 19M Repos Manual Filter Automatic Filter Query on
59. 59 19M Repos Manual Filter Automatic Filter Commits dh7h3 md8ja j287h dij873 dj827h os837 82uan 28a08 2ja82 d0hk0 j29yd a7jf9 aio92 hnna2
60. 60 19M Repos Manual Filter Automatic Filter Commits dh7h3 md8ja j287h dij873 dj827h os837 82uan 28a08 2ja82 d0hk0 j29yd a7jf9 aio92 hnna2
61. 61 19M Repos Manual Filter Automatic Filter Commits dh7h3 md8ja j287h dij873 dj827h os837 82uan 28a08 2ja82 d0hk0 j29yd a7jf9 aio92 hnna2
62. 62 19M Repos Manual Filter Automatic Filter Commits dh7h3 md8ja j287h dij873 dj827h os837 82uan 28a08 2ja82 d0hk0 j29yd a7jf9 aio92 hnna2 All commits were, at least, double-checked!
63. 63 19M Repos Manual Filter Automatic Filter 371 Energy-Aware Commits Query on
64. Research Questions 64 • RQ1. What are the solutions that developers use to save energy in practice? • RQ2. What software quality attributes may be given precedence over energy consumption? • RQ3. How are energy-saving solutions distributed over the software stack? • RQ4. To what extent are software developers certain that their commits will save energy?
65. RQ1: Solutions 65 • Frequency and voltage scaling (50 occurrences) • Use power efficient library/device (45 occurrences) • Disabling features or devices (42 occurrences) • Energy bug fix (26 occurrences) • Low power idling (22 occurrences) • Timing out (16 occurrences)
66. RQ1: Solutions 66 • Frequency and voltage scaling (50 occurrences) • Use power efficient library/device (45 occurrences) • Disabling features or devices (42 occurrences) • Energy bug fix (26 occurrences) • Low power idling (22 occurrences) • Timing out (16 occurrences)
67. RQ2: Quality Attributes 67 • Correctness (7 occurrences) • Responsiveness (6 occurrences) • Performance (3 occurrences) • No actual power saving (3 occurrences) • Miscellaneous (3 occurrences)
68. 68 • Correctness (7 occurrences) • Responsiveness (6 occurrences) • Performance (3 occurrences) • No actual power saving (3 occurrences) • Miscellaneous (3 occurrences) RQ2: Quality Attributes
69. 69 • Responsiveness (6 occurrences) • Responsiveness (6 occurrences) • Performance (3 occurrences) • No actual power saving (3 occurrences) • Miscellaneous (3 occurrences) RQ2: Quality Attributes
70. RQ3: Software Stack 70
71. 71 88 Commits Application includes embedded applications, desktop application, and mobile applications. RQ3: Software Stack
72. 72 88 Commits Application includes embedded applications, desktop application, and mobile applications. 42 Embedded RQ3: Software Stack
73. 73 88 Commits Application includes embedded applications, desktop application, and mobile applications. 42 Embedded 21 Arduino RQ3: Software Stack
74. 74 50 Commits Libraries/Utilities include scripts and embedded libraries. RQ3: Software Stack
75. 75 50 Commits Libraries/Utilities include scripts and embedded libraries. 41 scripts RQ3: Software Stack
76. 76 142 Commits Operating System includes Kernels, Embedded Kernels, Drivers and Firmwares RQ3: Software Stack
77. 77 142 Commits Operating System includes Kernels, Embedded Kernels, Drivers and Firmwares 69 Kernels RQ3: Software Stack
78. 78 142 Commits Operating System includes Kernels, Embedded Kernels, Drivers and Firmwares 69 — OS Kernel 54 — Drivers RQ3: Software Stack
79. RQ4: Certain 79 “Hesitating” words • seem • might • doubt • could • hope • attempt • supposed • guess • likely
80. RQ4: Certain 80 “Hesitating” words • seem • might • doubt • could • hope • attempt • supposed • guess • likely 18 hesitations!
81. RQ4: Certain 81 18 Reverted Commits
82. RQ4: Certain 82 18 Reverted Commits 8/18 reverts the power efficient work queue!
83. RQ4: Certain 83 18 Reverted Commits 8/18 reverts the power efficient work queue! There is no silver bullet! Fred Brooks
84. 2016 84 ICSME’16 Bad programmers worry about the code. Good programmers worry about data structures and their relationships. Linus Tolvards
85. In case you are a Java programmer… 85
86. 86 List<Object> lists = …; • ArrayList • LinkedList
87. 87 List<Object> lists = new ArrayList<>();
88. 88 List<Object> lists = new ArrayList<>(); Thread
89. 89 • ArrayList • LinkedList • Vector • Collections.synchronizedList() • CopyOnWriteArrayList List<Object> lists = …; Non Thread-Safe Thread-Safe
90. 90 List<Object> lists = new Vector<>();
91. 91 List<Object> lists = new Vector<>(); Thread
92. 92 Thread Thread-safe! List<Object> lists = new Vector<>();
93. 93 Thread { } … List<Object> lists = new Vector<>();
94. 94 Thread { } … List<Object> lists = new Vector<>();
95. 95 List<Object> lists = new CopyOnWriteArrayList<>();
96. 96 Thread Thread-safe! List<Object> lists = new CopyOnWriteArrayList<>();
97. List<Object> lists = new CopyOnWriteArrayList<>(); 97 { } Thread Thread-safe!
98. 98 List<Object> lists = …; • ArrayList • LinkedList • Vector • Collections.synchronizedList() • CopyOnWriteArrayList Non Thread-Safe Thread-Safe
99. 99 List<Object> lists = …; Set<Objects> sets = …; Map<Object, Object> maps = …;
100. 100 List<Object> lists = …; Set<Objects> sets = …; Map<Object, Object> maps = …;
101. 101 List<Object> lists = …; Set<Objects> sets = …; Map<Object, Object> maps = …;
102. 16 Collections 102 List ArrayList Vector Collections.syncList() CopyOnWriteArrayList Set LinkedHashSet Collections.syncSet() CopyOnWriteArraySet ConcurrentSkipListSet ConcurrentHashSet ConcurrentHashSetV8 Map LinkedHashMap Hashtable Collections.syncMap() ConcurrentSkipListMap ConcurrentHashMap ConcurrentHashMapV8
103. 16 Collections 103 List ArrayList Vector Collections.syncList() CopyOnWriteArrayList Set LinkedHashSet Collections.syncSet() CopyOnWriteArraySet ConcurrentSkipListSet ConcurrentHashSet ConcurrentHashSetV8 Map LinkedHashMap Hashtable Collections.syncMap() ConcurrentSkipListMap ConcurrentHashMap ConcurrentHashMapV8 Non thread-safe Thread-safe
104. 16 Collections 104 List ArrayList Vector Collections.syncList() CopyOnWriteArrayList Set LinkedHashSet Collections.syncSet() CopyOnWriteArraySet ConcurrentSkipListSet ConcurrentHashSet ConcurrentHashSetV8 Map LinkedHashMap Hashtable Collections.syncMap() ConcurrentSkipListMap ConcurrentHashMap ConcurrentHashMapV8 Java 7 Java 8
105. 16 Collections 105 List ArrayList Vector Collections.syncList() CopyOnWriteArrayList Set LinkedHashSet Collections.syncSet() CopyOnWriteArraySet ConcurrentSkipListSet ConcurrentHashSet ConcurrentHashSetV8 Map LinkedHashMap Hashtable Collections.syncMap() ConcurrentSkipListMap ConcurrentHashMap ConcurrentHashMapV8 x 3 Operations Traversal Insertion Removal
106. 2 Environments AMD CPU: A 2×16-core, running Debian, 2.4 GHz, 64GB of memory, JDK version 1.7 .0 11, build 21. 106 Intel CPU: A 2×8-core (32-cores w/ hyper-threading), running Debian, 2.60GHz, with 64GB of memory, JDK version 1.7 .0 71, build 14.
107. 2 Environments AMD CPU: A 2×16-core, running Debian, 2.4 GHz, 64GB of memory, JDK version 1.7 .0 11, build 21. 107 Intel CPU: A 2×8-core (32-cores w/ hyper-threading), running Debian, 2.60GHz, with 64GB of memory, JDK version 1.7 .0 71, build 14. Hardware-based energy measurement Software-based energy measurement
108. 108 Traversal Insertion Removal Intel CPU AMD CPU Traversal Insertion Removal Lists
109. 109 ArrayList Vector Lists Collections.synchronizedList()
110. 110 Lists ArrayList Vector Collections.synchronizedList() Energy (Joules)
111. 111 Lists ArrayList Vector Energy (Joules) Power (Watts) Collections.synchronizedList()
112. 112 Lists ArrayList Vector Energy (Joules) Power (Watts) Collections.synchronizedList()
113. Traversal Insertion Removal Intel CPU AMD CPU 113 113 Traversal Insertion Removal Lists
114. 114 Traversal Insertion Removal Intel CPU AMD CPU 114 Traversal Insertion Removal Lists
115. 115 Maps Intel CPU Traversal Insertion Removal AMD CPU Traversal Insertion Removal
116. 116 Maps Intel CPU Traversal Insertion Removal AMD CPU Traversal Insertion Removal
117. 117 Maps Intel CPU Traversal Insertion Removal AMD CPU Traversal Insertion Removal
118. 118 Maps Intel CPU Traversal Insertion Removal AMD CPU Traversal Insertion Removal
119. 119 Maps Intel CPU Traversal Insertion Removal AMD CPU Traversal Insertion Removal Less energy than the non thread-safe implementation!
120. Case Study 120
121. 121 Tomcat > A web server > More than 170K lines of Java code > More than 300 Hashtables
122. 122 Tomcat Xalan > Parses XML in HTML documents > More than 188K lines of Java code > More than 140 Hashtables > A web server > More than 170K lines of Java code > More than 300 Hashtables
123. 123 For each Hashtable instance, change it for a ConcurrentHashMap one. Do it again for ConcurrentHashMapV8 Task:
124. 124 Tomcat Hashtable to CHM: -12.21% Hashtable to CHM8: -17 .82%
125. 125 Tomcat Xalan Hashtable to CHM: -12.21% Hashtable to CHM8: -17 .82% Hashtable to CHM: -5.82% Hashtable to CHM8: -9.32%
126. 126 So, lets change all Hashtable instances to ConcurrentHashMap! Not so fast, young padawan…
127. 127 Hashtable ConcurrentHashMap Map 
 implements
128. 128 Hashtable ConcurrentHashMap Cloneable 
 Map 
 implements
129. 129 Hashtable ConcurrentHashMap Cloneable 
 Map 
 / / works fine implements Map<X,Y> obj = new Hashtable<>(); obj.clone();
130. 130 Hashtable ConcurrentHashMap Map<X,Y> obj = new Hashtable<>(); obj.clone(); Cloneable 
 Map 
 / / works fine / / compiler error implements Map<X,Y> obj = new ConcurrentHashMap<>(); obj.clone();
131. 131 Hashtable ConcurrentHashMap Cloneable 
 Map 
 / / works fine / / compiler error implements Danny Dig Opportunity for improving refactoring tools! Map<X,Y> obj = new Hashtable<>(); obj.clone(); Map<X,Y> obj = new ConcurrentHashMap<>(); obj.clone();
132. 2017 132 ASE’17 What are the most common bottlenecks in Java parallel computations?
133. 133 ForkJoin Projects Few representative OSS projects use ForkJoin
134. 134 + ForkJoin Projects
135. 135 actor
136. 136 actor actor
137. 137 actor
138. 138 actor FIFO (Centralized)
139. 139 actor FIFO (Centralized) 1s 2s 1s 92s 1s 1s
140. 140 actor FIFO (Centralized) 1s 2s 1s 92s 1s 1s actor Decentralized 1s 2s 1s 92s 1s 1s If a message is being executed when another message arrives, fork the second message
142. 142 #1. Copy on Fork #2. Copy on Join #3. Scattered Data #4. Exacting Intra-Task Synchronization #5. Bottleneck 6: Sleepy Workers
144. Practical Impact? 144
145. Practical Impact? 145 7/9 of projects that replied have accepted
146. 146 7/9 of projects that replied have accepted Practical Impact?
147. 147 7/9 of projects that replied have accepted Practical Impact?
150. Go to conferences. Talk to people!
151. Aim high!
152. Stay close to OSS communities!
153. Work on HARD stuff!
155. Dreams CACM’2017 Better tool support Books, cookbooks, guidelines Refactoring Testing Debugging Visualization Estimation ….
156. The Last Five Years of Energy Consumption Research Gustavo Pinto My personal, biased view of @gustavopinto gpinto@ufpa.br