Final Presentation @ ACM Student Research Competition ICSE'17

1. Locating Energy Hotspots in
Source Code
Rui Pereira
Universidade do Minho & INESC Tec/HASLab
ACM SRC @ ICSE’17
ruipereira@di.uminho.pt
26/05/2017 greenlab.di.uminho.pt

2. Going Green
2
+ =

3. Global energy system is unsustainable
3

4. Green Computing
4
 Caught the attention of many companies allowing them to save:
“close to 50% of the energy costs of an organization can be attributed to
the IT departments”
- [PICMET, 2009]
“up to 90% of energy used by ICT hardware can be attributed to software”
- [The Greenhouse Gas Protocol Report, 2013]

5. Green Software
 Reducing energy consumption through software analysis and optimization
 Problem:
 How to analyze
 How to interpret
 How to improve
5

6. Green Software
 Problems (extended to programmers):
 How to analyze
 How to interpret
 How to improve
6
Mining questions about software energy consumption
- [MSR’14]
Integrated energy-directed test suite optimization
- [ISTA’14]
Seeds: A software engineer’s energy-optimization decision
support framework - [ICSE’14]

7. Energy vs. Power
7
 Power (w) – rate (or effort) at which that work is done
 Energy (J) – amount of work done
 Power can change constantly while Energy is the accumulation
Energy = Power x Seconds
Power
Energy
100W360,000 J = x 3,600s

8. SPELL - Spectrum-based Energy Leak Localization
8

9. Spectrum-based Fault Localization (SFL)
9

10. Spectrum-based Fault Localization (SFL)
10
Error vector
Oracle
Low
High
Energy
Energy
Energy
Energy
Problem: Energy cannot be defined as binary values

11. SPELL
11

12. Validation
greenlab.di.uminho.pt 12
 (RQ1) – Does SPELL indicate reliable information?
 (RQ2) – Do developers actually find SPELL information useful?

13. Empirical Validation
greenlab.di.uminho.pt
Optimizing energy leaks with developers
Industrial
Post-doc
PhD
SPELL
13

14. SPELL – initial results
14
 Developers found the information to be very useful
 Spent on average 43% less time
 Produced more energy efficient programs (26% less energy on average)
 In progress:
 SPELL vs. Profilers (Looks promising!)
 Find Open-source Software and notify developers of possible hotspots and see what they say

15. SPELL – Plugin?
15
Code flagging Sunburst representation Suggestions?

16. SPELL – Accomplished Goals
16

17. Conclusions
17
Find more info @ our page: http://greenlab.di.uminho.pt/
Language independent
 Multi level analysis
 Multi hardware component analysis
 Points to hot spots in source code

18. 18
Thank you

19. Locating Energy Hotspots in
Source Code
Rui Pereira
Universidade do Minho & INESC Tec/HASLab
ACM SRC @ ICSE’17
ruipereira@di.uminho.pt
26/05/2017 greenlab.di.uminho.pt

20. Cheat slides
20

21. Is faster greener?
21
Understanding energy behaviors of thread management constructs
- [OOPSLA’14]
Energy-aware software: Challenges, opportunities and strategies
- [Journal of Computational Science’13]
Haskell in green land: Analyzing the energy behavior of a purely
functional language - [SANER’16]
An experimental survey of energy management across the stack
- [OOPSLA’14]
Power and energy implications of the number of threads used
- [PPMG’15]
Using the greenup, powerup, and speedup metrics to evaluate
software energy efficiency - [IGSC’15]
slower
Is faster greener?

22. 22
What is greener?

23. jRAPL/RAPL
23

24. What is jRAPL?
 Framework for profiling Java programs running on RAPL
24
What is RAPL?
 Running Average Power Limit
 Designed by intel for i5/i7 architectures (SandyBridge, IvyBridge, Haswell, etc)
 Monitors energy consumption info for Machine-Specific Registers (MSRs)
 Allows very precise and fine-grain measurements through function calls
 DRAM/GPU
 CPU
 Package

25. How do we use jRAPL?
25
Energy?
t0
t∞
a = getEnergy()
b = getEnergy()
b - a
 http://kliu20.github.io/jRAPL/

26. SPELL
26

27. 27 Example

28. 28 Example
ϕ - Component Category Similarty
37 + 61 + 0 + 42
1 + 2 + 0 + 1
75 + 102 + 0 + 34
140
4
211
=

29. 29 Example
ϕ - Component Category Similarty
Component c1
min(37,140)+min(38,166)+min(36,129)+min(37,164)+min(39,209)
max(37,140)+max(38,166)+max(36,129)+max(37,164)+max(39,209)
= 0.2314
𝝰1 =
𝝰2 =
𝝰3 =
min(1,4)+min(3,7)+min(1,3)+min(3,7)+min(2,11)
max(1,4)+max(3,7)+max(1,3)+max(3,7)+max(2,11)
= 0.3125
min(75,211)+min(77,259)+min(73,218)+min(74,222)+min(75,295)
max(75,211)+max(77,259)+max(73,218)+max(74,222)+max(75,295)
= 0.3104

30. 30 Example
𝜓 - Global Similarity

31. 31 Example
𝜓 - Global Similarity
(37 x 0.34) x 1 x 75 = 943.5
Component c1 Test t1
(61 x 0.34) x 2 x 102 = 4230.96
Component c2Test t1
(0 x 0) x 0 x 0 = 0
Component c3 Test t1
(42 x 0.34) x 1 x 34 = 485.52
Component c3 Test t1
= 5659.98

32. 32 Example
𝜓 - Global Similarity
Component c1
min(943.5,5659.98)+...+min(1989,14310.6)
max(943.5,5659.98)+...+max(1989,14310.6)
= 0.2466
𝜓1 =

33. Example