1. Power to the People: Leveraging Human Physiological Traits for Microprocessor Frequency Control
Alex Shye, Yan Pan, Ben Scholbrock, J. Scott Miller,
Gokhan Memik, Peter A. Dinda, Robert P. Dick
Northwestern University, EECS
ESP Project:
International Symposium on Microarchitecture, November 11, Lake Como, Italy.

2. Summary Claim: Any optimization ultimately exists to satisfy the user
Observation: Architectures largely ignore the individual user
Summary of Findings/Contributions
Make a case for adding biometric input devices to future architectures
Show that biometric devices can be used to indicate changes in user satisfaction as performance is altered
Demonstrate that these devices can be leveraged for user-aware optimization
Our goal is to attack this problem
11/11/08
International Symposium on Microarchitecture (MICRO-41), Lake Como, Italy

3. Why care about the user? 1 2 3 User-centric applications
Architectural trade-offs
exposed to the user
1) Iphone inherently interactive
Nintendo Wii : user experience over pure computational power
3) Mention ISCA paper
We are proposing a change in how humans and computers interact 3
Optimization opportunity
User variation = optimization potential
11/11/08
International Symposium on Microarchitecture (MICRO-41), Lake Como, Italy

4. Typical User Interaction
User Direction (from keyboard, mouse,etc.)
Output (from display,speakers,etc.)
1-sided interaction.
Note that computer has no information about the human.
Decisions to make affect perceived performance
11/11/08
International Symposium on Microarchitecture (MICRO-41), Lake Como, Italy

5. From the computer’s perspective
Performance Level ?
Problem: We can’t get user-related information that we’d like.
Explicit input possible, but not practical and may be annoying.
Without the appropriate information, it is difficult (if not
impossible) for the computer to take the user into account
11/11/08
International Symposium on Microarchitecture (MICRO-41), Lake Como, Italy

6. Our Goal 1 2 Provide computer user-related
Physiological traits (biometric inputs) 1
Provide computer user-related
information with biometric inputs
Performance Level
Informed 2
Leverage human physiological traits
for user-aware optimization
1) Input devices with sole purpose of getting user-related information
11/11/08
International Symposium on Microarchitecture (MICRO-41), Lake Como, Italy

7. Biometric Input Devices
Hypothesis: A change in human state due to changes in performance should be reflected by a change in physiological traits
We explore using three biometric devices:
Eye tracker
Galvanic skin response (GSR) sensor
Force sensors
11/11/08
International Symposium on Microarchitecture (MICRO-41), Lake Como, Italy

8. Eye Tracker Process video feed for: 2 measurements: PupilRadius
X-Y Coordinates of pupil on video
2 measurements:
PupilRadius
Mental workload [Iqbal CHI2005]
Perceptual changes [Einhauser NAS 2008]
Emotion processing [Partala JHCS 2003]
PupilMovement
Event Perception [Smith ETRA 2006]
11/11/08
International Symposium on Microarchitecture (MICRO-41), Lake Como, Italy

9. Galvanic Skin Response (GSR)
Conductance of skin
Reflects “fight-or-flight” response
Increases with engagement
Decreases with relaxation
11/11/08
International Symposium on Microarchitecture (MICRO-41), Lake Como, Italy

10. GSR Behavior GSR spikes with interest
DeltaGSR metric measures only the increases in GSR
If we want to measure GSR, we can’t just use the absolute value
11/11/08
International Symposium on Microarchitecture (MICRO-41), Lake Como, Italy

11. Force Sensors Piezoresistive Force Sensors Conductance α Force
MaxArrow
= Max(4 sensors)
We are not aware of previous work but the intuition behind this is that perhaps there will also be measureable behavioral differences as performance is changed.
11/11/08
International Symposium on Microarchitecture (MICRO-41), Lake Como, Italy

12. Sensor Selection They do not impede with the computer use
Require little effort to activate/mount
Can be easily integrated
Laptops contain integrated camera for eye tracking
Mouse/keyboard can be enhanced with GSR and force sensors
Power consumption negligible
“Cheap” extensions
We didn’t just select them randomly. There are several reasons we have selected these sensors.
11/11/08
International Symposium on Microarchitecture (MICRO-41), Lake Como, Italy

13. Sensor Metrics Four measurements: Sample at 30 Hz
PupilRadius, PupilMovement, DeltaGSR, MaxArrow
Sample at 30 Hz
Each second, compute three statistics:
Max, Mean, and Variance
Sensor metric = Statistic_Measurement
E.g., Max_MaxArrow and Mean_PupilRadius
30 Hz, each second we compute a sensor metric
Now that we have our sensors, we are ready to do user studies to understand the sensors
11/11/08
International Symposium on Microarchitecture (MICRO-41), Lake Como, Italy

14. User Study Setup IBM Thinkpad T61 Three user studies:
Intel Core 2 Duo CPU supporting Intel Speedstep (DVFS)
5 Frequencies (2.2Ghz Mhz)
Windows XP
Three user studies:
First two show that physiological traits change with performance
Third evaluates a system leveraging this information
Compare to an Adaptive DVFS scheme modeled after the Linux ondemand governor
Three interactive applications:
Need for Speed
Tetris Arena (third user study)
Microsoft Word (third user study)
Word is a non-CPU intensive application different than the games
11/11/08
International Symposium on Microarchitecture (MICRO-41), Lake Como, Italy

15. First User Study Goal: How:
Do human physiological traits change with changes in performance?
How:
14 users
Play Need for Speed
Drop performance to 600Mhz for 20 seconds
At same point in game every time
Better wording to explain “same point every time”
11/11/08
International Symposium on Microarchitecture (MICRO-41), Lake Como, Italy

16. Mean_PupilMovement Decrease of pupil movement across most users
Explain why we normalized; only change is important
Decrease of pupil movement across most users
11/11/08
International Symposium on Microarchitecture (MICRO-41), Lake Como, Italy

17. Max_MaxArrow Decrease in arrow pressure across most users
Surprising, and unexpected that ALL users decreased key force
Decrease in arrow pressure across most users
11/11/08
International Symposium on Microarchitecture (MICRO-41), Lake Como, Italy

18. Max_DeltaGSR Change varies among users
We attribute this change to the fact that users react differently to changes in performance.
Nevertheless, we notice a definite change in DeltaGSR
Change varies among users
Some get more aroused (irritated)
Some get less aroused (bored)
11/11/08
International Symposium on Microarchitecture (MICRO-41), Lake Como, Italy

19. Second User Study Goal: How:
Can changes in physiological traits be distinguished during game play?
Are the changes correlated to user satisfaction?
How:
20 users
Play Need for Speed
Randomly change to each of four other frequencies twice
First time, just collect sensor metrics
Second time, ask for user satisfaction rating: 1 (bad) – 5 (good)
With this information, we need a way to compare or interpret different sets of sensor readings
11/11/08
International Symposium on Microarchitecture (MICRO-41), Lake Como, Italy

20. Detecting/Interpreting Changes in Sensor Readings
“Good” sensor metric behavior
If user satisfaction same, sensor metrics should remain same
If user satisfaction different, sensor metrics should reflect this
We develop a T-test-based Similarity Metric
T-test distribution of sensor metric samples from different frequencies
High confidence indicates difference in user satisfaction
Low confidence indicates no change in user satisfaction
We now compute the t-test across the users and sensor metrics
11/11/08
International Symposium on Microarchitecture (MICRO-41), Lake Como, Italy

21. Using the T-test Similarity Metric
We adopt an
85% confidence
threshold
Always compare to the highest frequency; it is known to be good performance.
11/11/08
International Symposium on Microarchitecture (MICRO-41), Lake Como, Italy

22. Sensor Metrics vs. User Satisfaction
Success: T-test prediction matches change in user satisfaction
False Positive: T-test prediction falsely predicts change
False Negative: T-test prediction falsely predicts no change
Sensor Metric
Success Rate
False Positive
False Negative
Max_PupilRadius
70.2%
14.3%
15.5%
Max_MaxArrow
69.0%
13.1%
17.9%
Mean_MaxArrow
Mean_PupilRadius
67.9%
11.9%
20.2%
Mean_PupilMovement
57.1%
29.8%
Max_DeltaGSR
58.3%
9.5%
32.1%
Now that we have a metric, we can evaluate prediction via biometric inputs.
False positives are lost opportunity; false negatives are more problematic
False negatives higher than we would have liked…
11/11/08
International Symposium on Microarchitecture (MICRO-41), Lake Como, Italy

23. Using Biometrics for Optimization
We have shown that:
Human physiological traits do change with performance
We can use biometric readings to distinguish these changes
We construct PTP to leverage biometric readings
Physiological Traits-based Power-management
Power To the People
Built on top of Adaptive DVFS
Tests physiological traits to find a performance level comfortable for the user (settled frequency)
Uses settled frequency to set a ceiling for Adaptive DVF
Overview: we have shown that…
11/11/08
International Symposium on Microarchitecture (MICRO-41), Lake Como, Italy

24. PTP Learning Algorithm
Start at highest frequency
Successively test lower frequencies one by one
Each frequency test consists of three trials
One trial consists of:
20 seconds at highest frequency, 20 seconds at test frequency
Compute T-test for sensor metrics
Majority vote across sensors
Majority vote across trials
If a majority vote says OK, try next frequency
If majority vote predicts difference, go up one frequency and settle there
More details in the paper.
“test” frequency and NOT lowest frequency
cPTP described in paper
11/11/08
International Symposium on Microarchitecture (MICRO-41), Lake Como, Italy

25. PTP Learning Algorithm
Finish with “we settle at 1.6 Ghz”.
11/11/08
International Symposium on Microarchitecture (MICRO-41), Lake Como, Italy

26. Third User Study: PTP Evaluation
Goal:
Does PTP work?
How
Run the learning algorithm to find the settled frequency for the individual user
Run once with PTP at the settled frequency and once with the Adaptive scheme
Order is randomized
2.5 minutes each
Ask for user satisfaction rating from 1 (bad) – 5 (good)
Measure total system power for comparison
11/11/08
International Symposium on Microarchitecture (MICRO-41), Lake Como, Italy

27. Need for Speed Slightly decrease user satisfaction
Decrease, but the user satisfaction is still high
cPTP: ~6% total system power savings
Slightly decrease user satisfaction
18% total system power savings
11/11/08
International Symposium on Microarchitecture (MICRO-41), Lake Como, Italy

28. Tetris No change to user satisfaction 33% total system power savings
11/11/08
International Symposium on Microarchitecture (MICRO-41), Lake Como, Italy

29. Microsoft Word No change to user satisfaction
Not same interactiveness, and low CPU-utilization
Were able to settle at a considerably lower frequency so still able to understand user satisfaction
No change to user satisfaction
2% total system power savings
11/11/08
International Symposium on Microarchitecture (MICRO-41), Lake Como, Italy

30. Conclusion
Motivate new biometric input devices for future architectures
Eye tracker, GSR, and force sensors
Human physiological traits change with performance
Show biometric inputs can be used to indicate user satisfaction
Demonstrate PTP for user-aware power management
18% total system power savings across three applications
Little to no change in user satisfaction
11/11/08
International Symposium on Microarchitecture (MICRO-41), Lake Como, Italy

31. Thank you! Questions? Alex Shye
ESP: Empathic Systems Project
This concludes my talk. My contact information and the project website are below for reference. Thank you for listening and I would be happy to answer any questions.
11/11/08
International Symposium on Microarchitecture (MICRO-41), Lake Como, Italy

32. Sensors and User Satisfaction
Work on transition into this slide
Note that this is the average trends; for many, not linear, just average is
11/11/08
International Symposium on Microarchitecture (MICRO-41), Lake Como, Italy