PES: Proactive Event Scheduling for Responsive and Energy-Efficient Mobile Web Computing

ISCA 2019 Talk. Presented by Yu Feng

Transcript

1. PES: Proactive Event Scheduling for Responsive and Energy-Efficient Mobile Web Computing Yu Feng with Yuhao Zhu Department of Computer Science University of Rochester ISCA 2019
2. Ubiquity of Mobile devices 2
3. Ubiquity of Mobile devices 2
4. Ubiquity of Mobile devices 3
5. Ubiquity of Mobile devices 3 ▸2x times mobile users than desktop users
6. Ubiquity of Mobile devices 3 ▸2x times mobile users than desktop users ▸76% of population are using mobile devices daily
7. Mobile Applications are Event-Driven 4
8. Mobile Applications are Event-Driven 4
9. Mobile Applications are Event-Driven 4 Typing Interactions ISCA 2019
10. Mobile Applications are Event-Driven 4 Scrolling Typing Interactions ISCA 2019
11. Mobile Applications are Event-Driven 4 Scrolling Tapping Typing Interactions ISCA 2019
12. Mobile Applications are Event-Driven 4 Scrolling Tapping scrollstart scrollupdate scrollend tapdown Tapup Keypress Typing Interactions Events ISCA 2019
15. Event-Driven Execution Model 6 Event
16. Event-Driven Execution Model 6 touchmove onclick timer … Time Event Queue Push Event
17. Event-Driven Execution Model 7 touchmove onclick timer … Time Event Queue Fetch Runtime Scheduler Core
18. Event-Driven Execution Model 7 touchmove onclick timer … Time Event Queue Fetch Runtime Scheduler Core Scheduling knobs: DVFS settings
19. Inefficiency of Current Schedulers 8
20. Inefficiency of Current Schedulers 8 Time
21. Inefficiency of Current Schedulers 8 Time input 1 input 2 QoS Deadline 1 QoS Deadline 2
22. Inefficiency of Current Schedulers 8 Time OS Governor input 1 input 2 QoS Deadline 1 QoS Deadline 2
23. Inefficiency of Current Schedulers 8 Time OS Governor input 1 input 2 QoS Deadline 1 QoS Deadline 2 E1
24. Inefficiency of Current Schedulers 8 Time OS Governor input 1 input 2 QoS Deadline 1 QoS Deadline 2 E1 slack
25. Inefficiency of Current Schedulers 8 Time OS Governor input 1 input 2 QoS Deadline 1 QoS Deadline 2 E1 E2 slack
26. Inefficiency of Current Schedulers 9 Time input 1 input 2 QoS Deadline 1 QoS Deadline 2 E1 E2 OS Governor
27. Inefficiency of Current Schedulers 9 Time input 1 input 2 QoS Deadline 1 QoS Deadline 2 E1 E2 QoS-Aware OS Governor
28. Inefficiency of Current Schedulers 9 Time input 1 input 2 QoS Deadline 1 QoS Deadline 2 E1 E2 E1 QoS-Aware OS Governor
29. Inefficiency of Current Schedulers 9 Time input 1 input 2 QoS Deadline 1 QoS Deadline 2 E1 E2 E1 E2 QoS-Aware OS Governor
30. Inefficiency of Current Schedulers 9 Time input 1 input 2 QoS Deadline 1 QoS Deadline 2 E1 E2 E1 E2 QoS-Aware OS Governor ?
31. Inefficiency of Current Schedulers 10 Time input 1 input 2 QoS Deadline 1 QoS Deadline 2 E1 E2 E1 E2 QoS-Aware OS Governor
32. Inefficiency of Current Schedulers 10 Time Oracle input 1 input 2 QoS Deadline 1 QoS Deadline 2 E1 E2 E1 E2 QoS-Aware OS Governor
33. Inefficiency of Current Schedulers 10 Time Oracle input 1 input 2 QoS Deadline 1 QoS Deadline 2 E1 E2 E1 E2 E1 E2 QoS-Aware OS Governor
34. Inefficiency of Current Schedulers 11 Time E1 E2 E3 E4 E5 Schedule across events
35. Inefficiency of Current Schedulers 12 Time E1 E2 E3 E4 E5 Current events
36. Inefficiency of Current Schedulers 12 Time E1 E2 E3 E4 E5 Current events Future events
37. Proactive Event Scheduler 13
38. Proactive Event Scheduler Proactive Event Scheduler 13
39. Proactive Event Scheduler Proactive Event Scheduler 13 Prediction Web Program Analysis Machine Learning
40. Proactive Event Scheduler Proactive Event Scheduler 13 Prediction Web Program Analysis Machine Learning Scheduling Constraint Optimization
41. Proactive Event Scheduler Proactive Event Scheduler 13 Result reduce energy consumption while improving QoS Prediction Web Program Analysis Machine Learning Scheduling Constraint Optimization
42. Proactive Event Scheduler Proactive Event Scheduler 13 Result reduce energy consumption while improving QoS Prediction Web Program Analysis Machine Learning Scheduling Constraint Optimization
43. Event Sequence Learning 14 Prediction Model
44. Event Sequence Learning 14 Time E1 E2 E3 Event Sequence Prediction Model
45. Event Sequence Learning 14 Time E1 E2 E3 ? Event Sequence Prediction Model
46. Event Sequence Learning 14 Time E1 E2 E3 ? Event Sequence Prediction Model
47. Event Sequence Learning 14 Time E1 E2 E3 ? Event Sequence E4 Prediction Model
48. Event Sequence Learning 14 Time E1 E2 E3 ? Event Sequence E4 Prediction Model
49. Recurrent Prediction 15 Time E1 E2 E3 E4 Event Sequence Prediction Model
50. Recurrent Prediction 15 Time E1 E2 E3 E4 Event Sequence Prediction Model
51. Recurrent Prediction 15 Time E1 E2 E3 E4 Event Sequence E5 Prediction Model
52. Recurrent Prediction 15 Time E1 E2 E3 E4 Event Sequence E5 Prediction Model
53. Recurrent Prediction 15 Time E1 E2 E3 E4 Event Sequence E5 … Prediction Model
54. Prediction Model 16 Prediction Model
55. Prediction Model 17 Prediction Model
56. Prediction Model 17 Prediction Model The distance of click The number of scrolls The number of navigations Features encoding past interactions …
57. Prediction Model 18 Features encoding past interactions ln( p 1 − p ) = x β Prediction Model
58. Prediction Model 19 Features encoding past interactions Click ScrollUp ScrollDown ZoomIn ZoomOut … Prediction Model 0.10 0.12 0.58 0.07 0.02
59. Prediction Model 19 Features encoding past interactions Click ScrollUp ScrollDown ZoomIn ZoomOut … Prediction Model ✔
60. Prediction Model 19 Features encoding past interactions Click ScrollUp ScrollDown ZoomIn ZoomOut … Prediction Model All Event Types
61. Prediction Model 19 Features encoding past interactions Click ScrollUp ScrollDown ZoomIn ZoomOut … Prediction Model Filter Events ?
62. 20 Program Analysis
63. 20 Program Analysis
64. 20 Program Analysis
65. 20 Program Analysis Viewport
66. 20 <collapsible> <href> <html> <body> <div> … <div> <div> <href> <href> … Program Analysis DOM Tree Viewport
67. 20 <collapsible> <href> <html> <body> <div> … <div> <div> <href> <href> … Viewport Program Analysis DOM Tree Viewport
68. 20 <collapsible> <href> <html> <body> <div> … <div> <div> <href> <href> … Viewport Program Analysis DOM Tree Viewport
69. 20 <collapsible> <href> <html> <body> <div> … <div> <div> <href> <href> … Viewport Program Analysis DOM Tree Viewport
70. 20 <collapsible> <href> <html> <body> <div> … <div> <div> <href> <href> … Viewport Program Analysis DOM Tree Viewport
71. Include Program Analysis Features 21 Features encoding past interactions Click ScrollUp ScrollDown ZoomIn ZoomOut … Prediction Model
72. Overview of Predictor 22 Past Event Statistics Current Application State Click ScrollUp ScrollDown ZoomIn ZoomOut … + Prediction Model
73. Overview of Predictor 22 Past Event Statistics Current Application State Click ScrollUp ScrollDown ZoomIn ZoomOut … + Prediction Model Recurrent
74. Overview of Predictor 22
75. Overview of Predictor 22 High Accuracy, low overhead!
76. Proactive Event Scheduler Proactive Event Scheduler 23 Result reduce energy consumption meanwhile improve QoS Prediction Web Program Analysis Machine Learning Scheduling Constraint Optimization
77. Proactive Event Scheduler Proactive Event Scheduler 23 Result reduce energy consumption meanwhile improve QoS Prediction Web Program Analysis Machine Learning Scheduling Constraint Optimization
78. Scheduling events 24 E1 E2 E3 Time From predictor
79. Scheduling events 24 E1 E2 E3 Time From predictor Goal: minimize total energy while meeting deadlines
80. Problem Formulation 25 E1 E2 E3 Time ▸ Scheduling Problem → Constrained Optimization.
81. Objective: Problem Formulation 25 E1 E2 E3 Time ▸ Scheduling Problem → Constrained Optimization. N ∑ i Min. Energy (i)
82. Objective: Problem Formulation 25 E1 E2 E3 Time △Texe 1 △Texe 2 △Texe 3 ▸ Scheduling Problem → Constrained Optimization. N ∑ i △Texe (i) x Min.
83. Objective: Problem Formulation 25 E1 E2 E3 Time △Texe 1 △Texe 2 △Texe 3 ▸ Scheduling Problem → Constrained Optimization. N ∑ i △Texe (i) x Power (i) Min.
84. Objective: Constraints: Problem Formulation 25 E1 E2 E3 Time △Texe 1 △Texe 2 △Texe 3 ▸ Scheduling Problem → Constrained Optimization. N ∑ i △Texe (i) x Power (i) Min.
85. Objective: Constraints: Problem Formulation 25 E1 E2 E3 Time △Texe 1 △Texe 2 △Texe 3 ▸ Scheduling Problem → Constrained Optimization. Order: ≤ Tend (i) Tstart (i+1) N ∑ i △Texe (i) x Power (i) Min.
86. Objective: Constraints: Problem Formulation 25 E1 E2 E3 Time Tstart 1 Tstart 2 Tstart 3 TQoS 1 TQoS 2 TQoS 3 △Texe 1 △Texe 2 △Texe 3 ▸ Scheduling Problem → Constrained Optimization. Order: ≤ Tend (i) Tstart (i+1) Deadline: ≤ Tstart (i) △Texe (i) TQoS (i) + N ∑ i △Texe (i) x Power (i) Min.
87. Objective: Constraints: Problem Formulation 25 E1 E2 E3 Time Tstart 1 Tstart 2 Tstart 3 TQoS 1 TQoS 2 TQoS 3 △Texe 1 △Texe 2 △Texe 3 ▸ Scheduling Problem → Constrained Optimization. Order: ≤ Tend (i) Tstart (i+1) Deadline: ≤ Tstart (i) △Texe (i) TQoS (i) + Scheduling knobs: DVFS settings for each event N ∑ i △Texe (i) x Power (i) Min.
88. Objective: Constraints: Problem Formulation 26 ▸ Scheduling Problem → Constrained Optimization. N ∑ i △Texe (i) x Power (i) Min. Order: ≤ Tend (i) Tstart (i+1) Deadline: ≤ Tstart (i) △Texe (i) TQoS (i) +
89. Each Event: Objective: Constraints: Problem Formulation 26 ▸ Scheduling Problem → Constrained Optimization. N ∑ i △Texe (i) x Power (i) Min. △Texe (i) = Order: ≤ Tend (i) Tstart (i+1) Deadline: ≤ Tstart (i) △Texe (i) TQoS (i) +
90. Each Event: Objective: Constraints: Problem Formulation 26 ▸ Scheduling Problem → Constrained Optimization. N ∑ i △Texe (i) x Power (i) Min. △Texe (i) = Tmemory + Order: ≤ Tend (i) Tstart (i+1) Deadline: ≤ Tstart (i) △Texe (i) TQoS (i) +
91. Each Event: Objective: Constraints: Problem Formulation 26 ▸ Scheduling Problem → Constrained Optimization. N ∑ i △Texe (i) x Power (i) Min. Tcpu △Texe (i) = Tmemory + Order: ≤ Tend (i) Tstart (i+1) Deadline: ≤ Tstart (i) △Texe (i) TQoS (i) +
92. Each Event: Objective: Constraints: Problem Formulation 26 ▸ Scheduling Problem → Constrained Optimization. N ∑ i △Texe (i) x Power (i) Min. △Texe (i) = Tmemory + Ncycles / f (i) Order: ≤ Tend (i) Tstart (i+1) Deadline: ≤ Tstart (i) △Texe (i) TQoS (i) +
93. Each Event: Objective: Constraints: Problem Formulation 26 ▸ Scheduling Problem → Constrained Optimization. N ∑ i △Texe (i) x Power (i) Min. △Texe (i) = Tmemory + Constants Ncycles / f (i) Order: ≤ Tend (i) Tstart (i+1) Deadline: ≤ Tstart (i) △Texe (i) TQoS (i) +
94. Each Event: Objective: Constraints: Problem Formulation 26 ▸ Scheduling Problem → Constrained Optimization. N ∑ i △Texe (i) x Power (i) Min. △Texe (i) = Tmemory + Constants Offline profile Ncycles / f (i) Order: ≤ Tend (i) Tstart (i+1) Deadline: ≤ Tstart (i) △Texe (i) TQoS (i) +
95. Each Event: Objective: Constraints: Problem Formulation 26 ▸ Scheduling Problem → Constrained Optimization. N ∑ i △Texe (i) x Min. △Texe (i) = Tmemory + Constants Offline profile Ncycles / f (i) Order: ≤ Tend (i) Tstart (i+1) Deadline: ≤ Tstart (i) △Texe (i) TQoS (i) + Pmap (i)
96. Each Event: Objective: Constraints: Problem Formulation 27 ▸ Scheduling Problem → Constrained Optimization. N ∑ i △Texe (i) x Pmap (i) Min. Order: ≤ Tend (i) Tstart (i+1) Deadline: ≤ Tstart (i) △Texe (i) TQoS (i) +
97. Each Event: Objective: Constraints: Problem Formulation 27 ▸ Scheduling Problem → Constrained Optimization. N ∑ i △Texe (i) x Pmap (i) Min. Order: ≤ Tend (i) Tstart (i+1) Deadline: ≤ Tstart (i) △Texe (i) TQoS (i) + { Ncycles / f (j) } △Texe (i) =Tmemory +∑ * ⍺ (i, j) ⍺ (i, j) in {0,1}
98. Each Event: Objective: Constraints: Problem Formulation 27 ▸ Scheduling Problem → Constrained Optimization. N ∑ i △Texe (i) x Pmap (i) Min. Order: ≤ Tend (i) Tstart (i+1) Deadline: ≤ Tstart (i) △Texe (i) TQoS (i) + { Ncycles / f (j) } △Texe (i) =Tmemory +∑ * ⍺ (i, j) Pmap (i) = Freq2Power (j) * ⍺ (i, j) ∑
99. Each Event: Objective: Constraints: Problem Formulation 27 ▸ Scheduling Problem → Constrained Optimization. N ∑ i △Texe (i) x Pmap (i) Min. Order: ≤ Tend (i) Tstart (i+1) Deadline: ≤ Tstart (i) △Texe (i) TQoS (i) + { Ncycles / f (j) } △Texe (i) =Tmemory +∑ * ⍺ (i, j) Pmap (i) = Freq2Power (j) * ⍺ (i, j) ∑ 1 = ⍺ (i, j) With the constraints: ∑
100. Each Event: Objective: Constraints: Problem Formulation 27 ▸ Scheduling Problem → Constrained Optimization. N ∑ i △Texe (i) x Pmap (i) Min. Order: ≤ Tend (i) Tstart (i+1) Deadline: ≤ Tstart (i) △Texe (i) TQoS (i) + { Ncycles / f (j) } △Texe (i) =Tmemory +∑ * ⍺ (i, j) Pmap (i) = Freq2Power (j) * ⍺ (i, j) ∑ 1 = ⍺ (i, j) With the constraints: ∑ Only one setting is chosen
101. Each Event: Objective: Constraints: Problem Formulation 27 ▸ Scheduling Problem → Constrained Optimization. N ∑ i △Texe (i) x Pmap (i) Min. Order: ≤ Tend (i) Tstart (i+1) Deadline: ≤ Tstart (i) △Texe (i) TQoS (i) + { Ncycles / f (j) } △Texe (i) =Tmemory +∑ * ⍺ (i, j) Pmap (i) = Freq2Power (j) * ⍺ (i, j) ∑ 1 = ⍺ (i, j) With the constraints: ∑
102. Each Event: Objective: Constraints: Problem Formulation 27 ▸ Scheduling Problem → Constrained Optimization. N ∑ i △Texe (i) x Pmap (i) Min. Order: ≤ Tend (i) Tstart (i+1) Deadline: ≤ Tstart (i) △Texe (i) TQoS (i) + { Ncycles / f (j) } △Texe (i) =Tmemory +∑ * ⍺ (i, j) Integer Linear Programing! Pmap (i) = Freq2Power (j) * ⍺ (i, j) ∑ 1 = ⍺ (i, j) With the constraints: ∑
103. Putting Things Together 28 Web Application Hardware Architecture
104. PES Putting Things Together 28 Web Application Hardware Architecture
105. PES Putting Things Together 28 Web Application Hardware Architecture Predictor Events
106. PES Putting Things Together 28 Web Application Hardware Architecture Predictor Events Scheduler Predictions
107. PES Putting Things Together 28 Web Application Hardware Architecture Predictor Events Speculative Schedules Scheduler Predictions
108. PES Putting Things Together 28 Web Application Hardware Architecture Predictor Controller Events Speculative Schedules Uncommitted Results Scheduler Predictions
109. PES Putting Things Together 28 Web Application Hardware Architecture Predictor Controller Events Speculative Schedules Uncommitted Results Scheduler Predictions
110. PES Putting Things Together 28 Web Application Hardware Architecture Predictor Controller Events Speculative Schedules Commit Uncommitted Results Scheduler Predictions
111. PES Putting Things Together 28 Web Application Hardware Architecture Predictor Controller Events Speculative Schedules Commit Uncommitted Results Recover Scheduler Predictions
112. Proactive Event Scheduler Proactive Event Scheduler 29 Result reduce energy consumption meanwhile improve QoS Prediction Web Program Analysis Machine Learning Scheduling Constraint Optimization
113. Proactive Event Scheduler Proactive Event Scheduler 29 Result reduce energy consumption meanwhile improve QoS Prediction Web Program Analysis Machine Learning Scheduling Constraint Optimization
114. Experimental Setup 30
115. Experimental Setup 30 Implemented our framework in Chromium on top of Android system
116. ODORID XU+E development board, which contains an Exynos 5410 SoC Experimental Setup 30 Implemented our framework in Chromium on top of Android system
117. ODORID XU+E development board, which contains an Exynos 5410 SoC Experimental Setup 30 Implemented our framework in Chromium on top of Android system UI-level record and replay for reproducibility. [ISPASS’15]
118. Evaluation ▸Baseline Mechanisms ▹Interactive governor (Interactive) — Android default ▹EBS: a state-of-the-art QoS-aware scheduler ▹Oracle: optimal scheduler 31 31 Time Oracle input 1 input 2 QoS Deadline 1 QoS Deadline 2 E1 E2 E1 E2 E1 E2 QoS-Aware OS Governor
119. Evaluation ▸Baseline Mechanisms ▹Interactive governor (Interactive) — Android default ▹EBS: a state-of-the-art QoS-aware scheduler ▹Oracle: optimal scheduler 31 ▸Metrics ▹Energy Consumption ▹QoS Violation 31
120. Evaluation ▸Baseline Mechanisms ▹Interactive governor (Interactive) — Android default ▹EBS: a state-of-the-art QoS-aware scheduler ▹Oracle: optimal scheduler 31 ▸Metrics ▹Energy Consumption ▹QoS Violation 31 ▸Applications ▹Top web applications (e.g., www.amazon.com)
121. Experimental Result 32
122. Experimental Result 32 Norn. Energy 0 0.25 0.5 0.75 1 163 msn slashdot youtube google amazon ebay sina espn bbc cnn twitter Interactive EBS PES Oracle
123. Experimental Result 32 Norn. Energy 0 0.25 0.5 0.75 1 163 msn slashdot youtube google amazon ebay sina espn bbc cnn twitter Interactive EBS PES Oracle
124. Experimental Result 32 Norn. Energy 0 0.25 0.5 0.75 1 163 msn slashdot youtube google amazon ebay sina espn bbc cnn twitter Interactive EBS PES Oracle
125. Experimental Result 32 Norn. Energy 0 0.25 0.5 0.75 1 163 msn slashdot youtube google amazon ebay sina espn bbc cnn twitter Interactive EBS PES Oracle
126. Experimental Result 32 Norn. Energy 0 0.25 0.5 0.75 1 163 msn slashdot youtube google amazon ebay sina espn bbc cnn twitter Interactive EBS PES Oracle
127. Experimental Result 32 Norn. Energy 0 0.25 0.5 0.75 1 163 msn slashdot youtube google amazon ebay sina espn bbc cnn twitter Interactive EBS PES Oracle QoS Violation 0 0.15 0.3 0.45 0.6 163 msn slashdot youtube google amazon ebay sina espn bbc cnn twitter Interactive EBS PES Oracle
128. Experimental Result 32 Norn. Energy 0 0.25 0.5 0.75 1 163 msn slashdot youtube google amazon ebay sina espn bbc cnn twitter Interactive EBS PES Oracle QoS Violation 0 0.15 0.3 0.45 0.6 163 msn slashdot youtube google amazon ebay sina espn bbc cnn twitter Interactive EBS PES Oracle
129. Experimental Result 32 Norn. Energy 0 0.25 0.5 0.75 1 163 msn slashdot youtube google amazon ebay sina espn bbc cnn twitter Interactive EBS PES Oracle QoS Violation 0 0.15 0.3 0.45 0.6 163 msn slashdot youtube google amazon ebay sina espn bbc cnn twitter Interactive EBS PES Oracle
130. Experimental Result 32 Norn. Energy 0 0.25 0.5 0.75 1 163 msn slashdot youtube google amazon ebay sina espn bbc cnn twitter Interactive EBS PES Oracle QoS Violation 0 0.15 0.3 0.45 0.6 163 msn slashdot youtube google amazon ebay sina espn bbc cnn twitter Interactive EBS PES Oracle 61% less QoS Violation and 26% of Energy Reduction
131. Conclusion 33 ‣ To better satisfy the user experience while minimizing the energy, coordinating across events is crucial for event-driven applications.
132. Conclusion 33 ‣ To better satisfy the user experience while minimizing the energy, coordinating across events is crucial for event-driven applications. ‣ PES combines statistic inference with application analysis on event prediction.
133. Conclusion 33 ‣ To better satisfy the user experience while minimizing the energy, coordinating across events is crucial for event-driven applications. ‣ PES achieves significant energy savings while reducing QoS violations. ‣ PES combines statistic inference with application analysis on event prediction.
134. PES can be applied to other event-driven applications!
135. Thanks