1. Jacob R. Lorch Microsoft Research
Using User Interface Event Information in Dynamic Voltage Scaling Algorithms
Jacob R. Lorch
Microsoft Research
Alan Jay Smith
University of California, Berkeley
11th International Symposium on Modeling, Analysis, and Simulation of Computer and Telecommunication Systems
October 13, 2003

2. Outline Background and motivation
What is dynamic voltage scaling?
Why are user interface events important in deciding what speed and voltage to run the CPU?
Questions about user interface events
What do we want to know about user interface events that can help design voltage scaling algorithms?
Methodology
Trace-driven analysis
Analysis results
What have we learned about the design of DVS algorithms?
Using User Interface Event Information in Dynamic Voltage Scaling Algorithms

3. Dynamic voltage scaling
The ability to quickly and efficiently change CPU supply voltage without rebooting
Appearing on commercial CPU’s (Transmeta, AMD, Intel)
Reducing voltage reduces energy consumption – E α V2
Lowering voltage necessitates lowering speed
When is lower energy worth the lower speed?
When we can meet deadlines at less than maximum speed
Using User Interface Event Information in Dynamic Voltage Scaling Algorithms

4. User interface tasks How can we detect tasks?
Real-time environment (uncommon)
Application instrumentation (impractical and unlikely)
Automatic inference
Inference from user interface events
Response time must be ms, giving deadline estimate
Goal: make sure all work the system does to satisfy a user interface event completes by the deadline • • •
Using User Interface Event Information in Dynamic Voltage Scaling Algorithms

5. Estimating task CPU need
Speed to use depends on expected CPU need
Classic approach
Run at CPU requirement divided by deadline
PACE approach
Speed up as task progresses so that if task is short we use low speed and voltage throughout
Optimal speed schedule depends on probability distribution of CPU need
Task Kc
Task Kc
Probability
Speed
Task Kc …
Cycles
Time
Using User Interface Event Information in Dynamic Voltage Scaling Algorithms

6. Questions
How much of the CPU’s time is spent on user interface events?
Is there an efficient heuristic to find when user interface tasks end?
How often does handling user interface events require waiting for I/O?
How quickly should we run the CPU to achieve good response time goals for user interface tasks?
Which DVS algorithms work best for user interface tasks?
Is there significant enough difference between different user interface events that a DVS algorithm should handle them differently?
Using User Interface Event Information in Dynamic Voltage Scaling Algorithms

7. Methodology
Using User Interface Event Information in Dynamic Voltage Scaling Algorithms

8. Trace-driven analysis
Typically, voltage scaling algorithms are evaluated on short benchmarks
Reflect artificial workload
We use traces to answer questions about the “real world.”
What applications get used and how often
How much CPU time real applications use
VTrace runs in background on Windows machines collecting info on threads, user interface events, etc.
Over 1,800 hours of activity on eight users’ machines
Using User Interface Event Information in Dynamic Voltage Scaling Algorithms

9. Identifying tasks
Event Thing that can cause a thread to do something new
Response Time thread runs between events
Dependent event Event triggered during response to other event
Trigger event Event not dependent on another
Task All responses transitively dependent on trigger event • A C • A D A C B C D D
Using User Interface Event Information in Dynamic Voltage Scaling Algorithms

10. Analysis Results
Using User Interface Event Information in Dynamic Voltage Scaling Algorithms

11. How much of the CPU’s time is spent on user interface events?
Conclusion: User interface tasks are only part of the answer to task inference. Timer messages may provide additional opportunities to detect tasks and deadlines.
Using User Interface Event Information in Dynamic Voltage Scaling Algorithms

12. Can we detect task completion efficiently?
Inferring task boundaries online is high-overhead
Must track what responses caused what events
Requires much interposition
Simpler approach: consider UI task complete when…
idle thread is running and no I/O is ongoing, or
next UI event arrives
Conclusion: This approach works well, except it may produce wrong results for certain applications.
Using User Interface Event Information in Dynamic Voltage Scaling Algorithms

13. How often do UI tasks wait for I/O?
DVS algorithms must treat I/O time specially
I/O power, time do not scale with CPU speed
Mouse clicks require I/O frequently, but other UI types do not
Conclusion: Anticipating I/O time is important for mouse click tasks, but less so for other UI tasks.
Using User Interface Event Information in Dynamic Voltage Scaling Algorithms

14. How fast should we run UI tasks?
CDF of task CPU needs reveals suggests speed needed before deadline
e.g., 99.5% of mouse moves would finish in 50 ms at 266 MHz
Conclusion: Mouse moves require minimal speed. Keystrokes need faster speeds, and mouse clicks require the most; no single speed works for all users, so dynamic evaluation is important.
Using User Interface Event Information in Dynamic Voltage Scaling Algorithms

15. Which DVS algorithm is best for UI tasks?
Keystroke tasks
Mouse click tasks
Comparison of Flat (constant speed); Past/Peg (use min, then max); Stepped (constant acceleration); PACE (optimal)
PACE always best, Stepped generally 2nd; PACE vs. Stepped: 7.3% less energy for keystrokes, 2.8% less for mouse clicks
Conclusion: Use PACE. If too much complexity or overhead, consider Stepped, especially for mouse clicks.
Using User Interface Event Information in Dynamic Voltage Scaling Algorithms

16. How best to estimate task CPU needs?
When a task starts, we estimate its CPU needs based on recent past tasks’ CPU needs: …
8.2 Kc
3.1 Kc
4.0 Kc
2.1 Kc
5.3 Kc
6.3 Kc
4.2 Kc
5.1 Kc
???
past
future
Task Kc
Task Kc
Probability
Speed
Task Kc
Task Kc …
Time
Cycles
PACE formula
Distribution
estimation
Using User Interface Event Information in Dynamic Voltage Scaling Algorithms

17. Separating tasks into groups
Tasks fall into different groups (e.g., enter keypresses in Excel vs. left mouse clicks in Word): …
8.2 Kc
3.1 Kc
4.0 Kc
2.1 Kc
5.3 Kc
6.3 Kc
4.2 Kc
5.1 Kc
???
past
future
Only basing prediction on similar tasks can improve prediction.
Task Kc
Task Kc
Probability
Speed
Task Kc
Task Kc …
Time
Cycles
PACE formula
Distribution
estimation
Using User Interface Event Information in Dynamic Voltage Scaling Algorithms

18. Overdoing task separation
If you make the distinctions too fine, there may be few tasks in the same group to match: …
8.2 Kc
3.1 Kc
4.0 Kc
2.1 Kc
5.3 Kc
6.3 Kc
4.2 Kc
5.1 Kc
???
past
future
Too many groups means you make your prediction based on little (or old) information.
Task Kc
Probability
Speed
Time
Cycles
PACE formula
Distribution
estimation
Using User Interface Event Information in Dynamic Voltage Scaling Algorithms

19. Does task grouping help or hurt prediction?
Tested four means of task grouping
N (None): Every task in same group
A (Application): Tasks in same app are in same group
C (Category): Tasks of same category (e.g., all “enter” keypresses) are in same group
CA (Category/Application): Separate group for each category and application combination
Measured average error in predicting CPU need, i.e., just the mean of the distribution
N worst, CA best – differences significant
Keystroke tasks
Conclusion: Fine task grouping helps prediction.
Using User Interface Event Information in Dynamic Voltage Scaling Algorithms

20. Does task grouping save energy?
Keystroke tasks
Mouse click tasks
To evaluate prediction of entire distribution, we simulated PACE
Results the same as for predicting mean: N worst, CA best
Conclusion: Fine task grouping helps prediction, which leads to energy savings. But, savings are minute, so fine grouping is only worthwhile if it adds virtually no complexity or overhead.
Using User Interface Event Information in Dynamic Voltage Scaling Algorithms

21. Conclusions
User interface tasks are a useful source of task information for DVS algorithms, but only account for 20.3% of CPU time
Our simple heuristic for task completion detection works well for most applications
Mouse click tasks wait for I/O frequently
Mouse movements require only minimal speed
PACE is the best DVS algorithm, but Stepped saves nearly as much energy and is simpler
Task grouping improves task CPU use prediction, but energy savings from such grouping are small
Using User Interface Event Information in Dynamic Voltage Scaling Algorithms