31
Arindam Mallik Jack Cosgrove Robert P. Dick Gokhan Memik Peter Dinda Northwestern University Department of Electrical Engineering and Computer Science Evanston, Illinois, USA ASPLOS March 3, 2008 Seattle, Washington, USA 1

Arindam Mallik Jack Cosgrove Robert P. Dick Gokhan Memik Peter Dinda Northwestern University Department of Electrical Engineering and Computer Science

Embed Size (px)

Citation preview

Page 1: Arindam Mallik Jack Cosgrove Robert P. Dick Gokhan Memik Peter Dinda Northwestern University Department of Electrical Engineering and Computer Science

Arindam MallikJack CosgroveRobert P. DickGokhan MemikPeter Dinda

Northwestern UniversityDepartment of Electrical Engineering and Computer ScienceEvanston, Illinois, USA

ASPLOS • March 3, 2008 • Seattle, Washington, USA 1

Page 2: Arindam Mallik Jack Cosgrove Robert P. Dick Gokhan Memik Peter Dinda Northwestern University Department of Electrical Engineering and Computer Science

Traditional performance metrics do not measure user-perceived performance well

Our performance metrics measure user-perceived performance better

PICSEL is a power management policy that uses our metrics to achieve system power improvements of up to 12.1% compared to existing policies

2

Page 3: Arindam Mallik Jack Cosgrove Robert P. Dick Gokhan Memik Peter Dinda Northwestern University Department of Electrical Engineering and Computer Science

3

CPU

Display Main Memory

Screenshot

Compare consecutive screenshots

Change frequency

Redrawscreen

Page 4: Arindam Mallik Jack Cosgrove Robert P. Dick Gokhan Memik Peter Dinda Northwestern University Department of Electrical Engineering and Computer Science

4

Display Main Memory

Screenshot

Compare consecutive screenshots

CPU

Change frequency

Redrawscreen

“The ultimate goal of a computer system is to satisfy the user”

Page 5: Arindam Mallik Jack Cosgrove Robert P. Dick Gokhan Memik Peter Dinda Northwestern University Department of Electrical Engineering and Computer Science

Power problem DVFS

System performance Traditional vs. user-perceived

PICSEL How it works Results

Conclusions

5

Page 6: Arindam Mallik Jack Cosgrove Robert P. Dick Gokhan Memik Peter Dinda Northwestern University Department of Electrical Engineering and Computer Science

Energy-hungry processors present three major problems: Higher energy consumption Shorter battery life Higher temperatures

6

Page 7: Arindam Mallik Jack Cosgrove Robert P. Dick Gokhan Memik Peter Dinda Northwestern University Department of Electrical Engineering and Computer Science

Dynamic voltage and frequency scaling (DVFS) addresses all three problems Trades off processor frequency for

energy savings Commonly used

Ideal DVFS policy: Find the lowest level of performance acceptable to the user to maximize power savings

7

Page 8: Arindam Mallik Jack Cosgrove Robert P. Dick Gokhan Memik Peter Dinda Northwestern University Department of Electrical Engineering and Computer Science

Human in loop is often rate-limiter

8

Output Devices

(kHz)User(Hz)

Processor(GHz)

Input Devices(kHz)

Page 9: Arindam Mallik Jack Cosgrove Robert P. Dick Gokhan Memik Peter Dinda Northwestern University Department of Electrical Engineering and Computer Science

Traditional performance metrics focus on processor performance “Close to metal”

9

Output Devices

User

Input Devices

Processor(IPS)

Page 10: Arindam Mallik Jack Cosgrove Robert P. Dick Gokhan Memik Peter Dinda Northwestern University Department of Electrical Engineering and Computer Science

10

User-perceived performance metrics focus on interface device performance “Close to flesh”Output Devices

(Display, Speakers)

User(N/A)

Input Devices(Mouse,

Keyboard)

Processor

Page 11: Arindam Mallik Jack Cosgrove Robert P. Dick Gokhan Memik Peter Dinda Northwestern University Department of Electrical Engineering and Computer Science

Use change in pixel intensities as metric for user-perceived performance

11

Page 12: Arindam Mallik Jack Cosgrove Robert P. Dick Gokhan Memik Peter Dinda Northwestern University Department of Electrical Engineering and Computer Science

PerceptionInformedCPU performanceScaling toExtend batteryLife

12

Page 13: Arindam Mallik Jack Cosgrove Robert P. Dick Gokhan Memik Peter Dinda Northwestern University Department of Electrical Engineering and Computer Science

Windows GDI Screenshot Capture contiguous area of screen Repeat periodically Compare RGB intensities across samples

13

Ri

Gi

Bi

Ri-1

Gi-1

Bi-1

- =

G Δ

B Δ

Cached

Page 14: Arindam Mallik Jack Cosgrove Robert P. Dick Gokhan Memik Peter Dinda Northwestern University Department of Electrical Engineering and Computer Science

Average Pixel Change (APC) APC = (RΔ + GΔ + BΔ) / 3 Averaged across all pixels Measures “slowness” of display

Rate of Average Pixel Change (APR) APR = (APCi – APCi-1)/(Ti – Ti-1) Measures “jitter” of display

14

Page 15: Arindam Mallik Jack Cosgrove Robert P. Dick Gokhan Memik Peter Dinda Northwestern University Department of Electrical Engineering and Computer Science

PICSEL uses <2% CPU utilizationCost of target applications is 50-

100% CPU utilization

15

Page 16: Arindam Mallik Jack Cosgrove Robert P. Dick Gokhan Memik Peter Dinda Northwestern University Department of Electrical Engineering and Computer Science

16

Page 17: Arindam Mallik Jack Cosgrove Robert P. Dick Gokhan Memik Peter Dinda Northwestern University Department of Electrical Engineering and Computer Science

17

APC

APR

Make a decision on these marks

Time

Increase frequency

“No change” band

“No change” band

Page 18: Arindam Mallik Jack Cosgrove Robert P. Dick Gokhan Memik Peter Dinda Northwestern University Department of Electrical Engineering and Computer Science

State Variables Adaptation Parameters

Processor frequency (f) Hysteresis factor (α)

APC in the last interval (μAPC) APC change threshold (ρ)

APR in the last interval (μAPR) APR change threshold (γ)

18

IF (APCinit - μAPC) < ρ ×(1-α) × APCinit

OR |APRinit - μAPR| < γ ×(1-α) × APRinit

Reduce f by one level

Reset α of the last level to 0.0

ELSE

Increase f by one level

Increment α by 0.1

Page 19: Arindam Mallik Jack Cosgrove Robert P. Dick Gokhan Memik Peter Dinda Northwestern University Department of Electrical Engineering and Computer Science

19

PICSEL Version Tinitialize

(sec)Tdecide

(sec)APC

ChangeAPR

ChangeHyst.Factor

Conservative PICSEL

(cPICSEL)10 7 0.05 0.15 0.0

Aggressive PICSEL

(aPICSEL)10 7 0.10 0.30 0.0All values chosen by authors after

testing using target applicationsToo long (243 days) to construct

ideal valuesUser evaluation “closed the loop”

Page 20: Arindam Mallik Jack Cosgrove Robert P. Dick Gokhan Memik Peter Dinda Northwestern University Department of Electrical Engineering and Computer Science

20

20 users Shockwave animation and DVD movie

play for 2 minutes FIFA game plays for 3.5 minutes Three randomly selected trials per

application One double-blind DVFS policy for each

trial User rates satisfaction from one (lowest)

to five (highest) after each trial

Page 21: Arindam Mallik Jack Cosgrove Robert P. Dick Gokhan Memik Peter Dinda Northwestern University Department of Electrical Engineering and Computer Science

21

Page 22: Arindam Mallik Jack Cosgrove Robert P. Dick Gokhan Memik Peter Dinda Northwestern University Department of Electrical Engineering and Computer Science

22

Page 23: Arindam Mallik Jack Cosgrove Robert P. Dick Gokhan Memik Peter Dinda Northwestern University Department of Electrical Engineering and Computer Science

23

Page 24: Arindam Mallik Jack Cosgrove Robert P. Dick Gokhan Memik Peter Dinda Northwestern University Department of Electrical Engineering and Computer Science

DVFS Policy

System Power

Improvement

Dynamic Power

Improvement

CPU Peak Temperat

ure Reduction

User Satisfacti

on(out of five)

aPICSEL 12.1% 18.2% 4.3C 3.65*

cPICSEL 7.1% 9.1% 1.7C 3.80**

Windows DVFS

Control Control Control 3.68

24

* Not Different with 95% confidence ** Different with 90% confidence

Page 25: Arindam Mallik Jack Cosgrove Robert P. Dick Gokhan Memik Peter Dinda Northwestern University Department of Electrical Engineering and Computer Science

DVFS Policy

System Power

Improvement

Dynamic Power

Improvement

CPU Peak Temperat

ure Reduction

User Satisfacti

on(out of five)

aPICSEL 12.1% 18.2% 4.3C 3.65*

cPICSEL 7.1% 9.1% 1.7C 3.80**Windows

DVFSControl Control Control 3.68

25

* Not Different with 95% confidence ** Different with 90% confidence

Page 26: Arindam Mallik Jack Cosgrove Robert P. Dick Gokhan Memik Peter Dinda Northwestern University Department of Electrical Engineering and Computer Science

26

Page 27: Arindam Mallik Jack Cosgrove Robert P. Dick Gokhan Memik Peter Dinda Northwestern University Department of Electrical Engineering and Computer Science

27

Perceived slowdown

Page 28: Arindam Mallik Jack Cosgrove Robert P. Dick Gokhan Memik Peter Dinda Northwestern University Department of Electrical Engineering and Computer Science

DVFS PolicyTotal Thermal

Emergencies during Game for All Users

aPICSEL 52

cPICSEL 51

Windows DVFS 59

28

User satisfaction is maximized by cPICSEL Frequency is high enough to deliver good

performance but not high enough to trigger thermal emergencies

Page 29: Arindam Mallik Jack Cosgrove Robert P. Dick Gokhan Memik Peter Dinda Northwestern University Department of Electrical Engineering and Computer Science

29

Display performance is a better metric for controlling DVFS than processor performance Existing processor performance-based

DVFS policies have slack that can be exploited

Cost of monitoring the display output is low

User satisfaction is the same or better

Page 30: Arindam Mallik Jack Cosgrove Robert P. Dick Gokhan Memik Peter Dinda Northwestern University Department of Electrical Engineering and Computer Science

Based on GUI events Gurun, S. and Krintz, C. 2005. AutoDVS: an Automatic,

General-purpose, Dynamic Clock Scheduling System for Hand-held Devices. In Proc. of the 5th ACM Int. Conf. on Embedded Software (EMSOFT’05), 218-226.

Based on application messages Flautner, K. and Mudge, T. 2002. Vertigo: Automatic

Performance-Setting for Linux. ACM SIGOPS Operating Systems Review 36, SI (Winter 2002), 105-116.

30

Page 31: Arindam Mallik Jack Cosgrove Robert P. Dick Gokhan Memik Peter Dinda Northwestern University Department of Electrical Engineering and Computer Science

Check out “Empathic Computer Architectures and Systems” at Wild and Crazy Ideas and visit

empathicsystems.org

for more user-centered systems research

31