User- and Process-Driven Dynamic Voltage and Frequency Scaling

1
User- and Process-DrivenDynamic Voltage and
Frequency Scaling

• Bin Lin
• Arindam Mallik
• Peter Dinda
• Gokhan Memik
• Robert Dick
• Empathic Systems Project
• Department of Electrical Engineering and Computer
  Science
• Northwestern University
• http//empathicsystems.org

2
Summary

• User-Driven Frequency Scaling (UDFS)
• Adjust CPU frequency according to satisfaction of
  individual user, exploiting variation in
  satisfaction among users
• Process-Driven Voltage Scaling (PDVS)
• Adjust voltage at frequency and temperature based
  on profile of the individual processor,
  exploiting variation among processors
• 50 power reduction on Windows tasks compared to
  Windows DVFS
• User study, measured system power

3
Outline

• Summary
• User-Driven frequency scaling (UDFS)
• User pessimism study
• UDFS2 algorithm
• Process-Driven voltage scaling (PDVS)
• PDVS profiling
• Evaluation of UDFS and PDVS
• Work in context
• Conclusion

4
User Satisfaction With CPU Frequency

• How satisfied are users of different applications
  at different clock frequencies?
• Earlier work suggests there will much variation
• User Study
• 8 users
• 3 frequencies Windows XP DVFS
• 3 typical Windows apps
• Presentation, Animation, Game
• Rate comfort on Leikert-like 1 to 10 scale

5

• Dramatic variation in user satisfaction for fixed
  frequencies
• And for DVFS
• Setting frequency for worse case user is
  pessimistic
• Setting it for an average user is also wrong

Presentation
Game
6
User-Driven Frequency Scaling (UDFS)

• Developed system to dynamically customize
  frequency to the individual user
• Key idea user feedback at run-time
• User presses irritation button (F11) as input
• But not too often!
• 2 very simple learning algorithms
• UDFS1, UDFS2

7
UDFS2 Algorithm

• Goal probe to find the lowest frequency level
  the user is comfortable at and stabilize there
• General idea react additively to both passage of
  time and user irritation events also update rate
  of reaction based on rate of user feedback
• UDSF1, in contrast, is an additive
  increase/multiplicative decrease algorithm,
  analogous to TCP Reno congestion control

8
UDFS2 Algorithm

• Each frequency level ri has interval ti
• ti 10 seconds, initially
• ri are frequencies in descending order
• User not irritated in current interval ti ?
• Switch frequency to ri1 , interval ti1
• Lets go slower in the next interval

9
UDFS2 Algorithm

• User irritated in current interval ti ?
• ti-1 a ti-1 (agt1)
• We should spend more time at the next highest
  frequency
• tk b tk " k k¹i-1 (blt1)
• We should spend less time at the other
  frequencies, including this one
• Switch frequency to ri-1 , interval ti-1
• a 2.5, b0.8 in our studies

10
UDFS2 in Action(User is playing a game)
Note Convergence
11
Outline

• Summary
• User-Driven frequency scaling (UDFS)
• User pessimism study
• UDFS2 algorithm
• Process-Driven voltage scaling (PDVS)
• PDVS profiling
• Evaluation of UDFS and PDVS
• Work in context
• Conclusion

12
CPU Voltage

• Minimum voltage needed for CPU stability is a
  function of frequency and temperature
• Claim function varies across individual parts
  for many interesting processors
• Hence, we can do better than the worst case
  (nominal) function
• P µ V2CF makes even small changes significant

13
Example
This processor can be run at lower voltages than
specifications indicate
2.13 GHz Pentium M-770 in a Thinkpad T43p
14
Process-Driven Voltage Scaling

• Profiling System (Boot CD)
• Workload generator
• Temperature monitor
• Frequency selector
• Voltage selector
• Watchdog timer
• For each frequency/temperature
• Modulate workload to maintain temperature
• Decrease voltage until watchdog timer fires and
  reboots machine
• Profiling system implemented by Nikolay
  Valtchanov and Matt Robben

Profile min_voltage(frequency, temperature)
Any DVFS Scheme
15
Outline

• Summary
• User-Driven frequency scaling (UDFS)
• User pessimism study
• UDFS2 algorithm
• Process-Driven voltage scaling (PDVS)
• PDVS profiling
• Evaluation of UDFS and PDVS
• Work in context
• Conclusion

16
Experimental Setup

• IBM Thinkpad T43p
• 2.13 GHz Pentium M-770
• 1 GB RAM
• Windows XP SP2
• UDFS system (or Windows DVFS) runs online,
  adjusting CPU frequency as user interacts with
  the system and applications
• PDVS effects and power measurements/analysis are
  done offline, using logs from UDFS and user
  traces
• System power measurement
• CPU dynamic power via simulation

17
Application Tasks

• Use Microsoft Powerpoint 2003 to replicate a
  presentation while listening to background music
  with Windows Media Player 10
• Watch 3D Shockwave animation with Microsoft
  Internet Explorer (locally stored animation)
• Play FIFA 2005 Soccer (first person shooter game)

18
Users

• 20 participants recruited from Northwestern
  population via IRB-approved methods
• Participants self-identified as Power User,
  Typical User, Beginner for each application,
  plus PCs and Windows
• A demographic mix

19
Study Process

• Fill out a questionnaire (2 minutes)
• Read a one page handout (2 minutes)
• Acclimatize to the performance of our machine by
  using the apps (5 minutes)
• Perform the following tasks for UDFS1
• Powerpoint (4 minutes)
• 3D Shockwave (4 minutes)
• FIFA game (8 minutes)
• Repeat previous for UDFS2

20
ResultsCompared to Windows DVFS Unless Otherwise
Noted

• Measured system power
• UDFS1, UDFS2, UDFS1PDVS, UDFS2PDVS, Windows
  DVFSPDVS
• Simulated CPU dynamic power
• UDFS1, UDFS2, UDFS1PDVS, UDFS2PDVS, Windows
  DVFSPDVS
• Measured mean and peak temperature
• UDFS1, UDFS2, UDFS1PDVS, UDFS2PDVS
• User irritation event rate (UDFS1,2)
• Multitasking study

21
Measured System Power(PowerPoint, improvement
over Windows DVFS)

• PDVS dominates gains for less interactive
  applications with little user-user variance

22
Measured System Power(FIFA, improvement over
Windows DVFS)

• UDFS contributes significant gains for more
  interactive applications

23
Measured Temperature(PowerPoint)

• Significant temperature reductions are possible
  using UDFS and PDVS

24
Measured Temperature(FIFA Game)

• Significant temperature reductions are possible
  using UDFS and PDVS

25
User Irritation Button Presses
First 4 minutes
Second 4 minutes

• Rate of user feedback can be low, and can
  decrease with time

26
Outline

• Summary
• User-Driven frequency scaling (UDFS)
• User pessimism study
• UDFS2 algorithm
• Process-Driven voltage scaling (PDVS)
• PDVS profiling
• Evaluation of UDFS and PDVS
• Work in context
• Conclusion

27
Work In Context of Power Management Efforts in
the Empathic Systems Project (empathicsystems.org)

• A. Gupta, B. Lin, P. Dinda, Measuring And
  Understanding User Comfort With Resource
  Borrowing, HPDC 2004
• Opportunity paper Identification of variance in
  user satisfaction with systems decisions
• P. Dinda, G. Memik, R. Dick, B. Lin, A. Mallik,
  A. Gupta, S. Rossoff, The User In Experimental
  Computer Systems Research, ExpCS 2007
• Position paper overview of goals of project and
  advocacy for user-driven work
• A. Mallik, J. Cosgrove, R. Dick, G. Memik, P.
  Dinda, PICSEL Measuring User-Perceived
  Performance to Control Dynamic Frequency Scaling,
  ASPLOS 2008
• Customize power management to user by observing
  output to him
• A. Shye, B. Ozisikyilmaz, A. Mallik, G. Memik, P.
  Dinda, R. Dick, A. Choudhary, Learning and
  Leveraging the Relationship between
  Architectural-level Measurements and Individual
  User Satisfaction, ISCA 2008
• Learn performance counter-gtuser satisfaction and
  use in power management
• A. Shye, Y. Pan, B. Scholbrock, J. S. Miller, G.
  Memik, P. Dinda, R. Dick, Power to the People
  Leveraging Human Physiological Traits to Control
  Microprocessor Frequency, MICRO 2008
• Measure user satisfaction using biometrics and
  use in power management
• This paper
• Direct user feedback for power management PDVS

28
Related Work

• DVFS in general
• Gochman, et al Intel Tech Journal, Broch, et al
  SOC03, and many more
• Dynamic Thermal Management
• Liu, et al IEEE JSSC-93, Brookes, et al
  WCED00, Crusoe, Intel Pentium-M, etc
• PDVS-related
• Teodorescu, et al ISCA08, Razor Ernst, et al,
  MICRO03, Dhar, et al ISLPED05, Intel Foxton,
  AutoDVS EMSOFT05, etc
• UDFS-related
• Lorch, et al UCB TR, Yan, et al DAC05,
  Vertigo OSDI02, Xu, et al EMSOFT05, Ranga, et
  al IEEE Computer-06, Anand, et al MOBICOMM03,
  Theocharous, et al Intel Tech Journal 06

29
Conclusion

• User-Driven Frequency Scaling (UDFS)
• Adjust CPU frequency according to satisfaction of
  individual user, exploiting variation in
  satisfaction among users
• Process-Driven Voltage Scaling (PDVS)
• Adjust voltage at frequency and temperature based
  on profile of the individual processor,
  exploiting variation among processors
• 50 power reduction on Windows tasks compared to
  Windows DVFS
• User study, measured system power
• Variation is opportunity

30
For MoreInformation

• Empathic Systems Project
• http//empathicsystems.org
• Prescience Lab
• http//presciencelab.org
• Peter Dinda
• http//pdinda.org

31
Measured System Power(Multitasking Study,
improvement over Windows DVFS))

• Multitasking appears to increase benefits
• Small additional study where user watched 3D
  animation while also listening to MP3s using
  Windows Media Player

32
Measured System Power(3D Shockwave,
improvement over Windows DVFS)

• UDFS contributes significant gains for more
  interactive applications

33
Measured Temperature(3D Shockwave)

• Significant temperature reductions are possible
  using UDFS and PDVS