Mobile Grid Major Area Examination

1
A Run-Time Feedback Based Energy Estimation Model
for Embedded SystemsSelim Gürün and Chandra
KrintzDepartment of Computer Science, U.C. Santa
Barbara, USA gurun,ckrintz_at_cs.ucsb.edu
Why power profiling?
Develop a power power model offline and update it
continuously to capture all behavioral dynamics
Our Approach
Power-aware methods divide task execution into
operations, and prepare an execution plan for each
Goals Run-time efficiency Accuracy Fine-grain
prediction
Baseline Local Reduced Remote Reduced

• Collect energy consumption every 10 million
  instructions
• Explore impact of measurement error
• Inject uniformly distributed random error
• Evaluate model size, update frequency and
  dynamicity

Measurement Precision
Avg. Error Rate
Knowing future energy cost of operations requires
profiling them at run-time
Execution Cost
Operation Unit of execution (context
dependent) E.g. In interactive tasks, smallest
user-visible unit of execution. Ideally A
Method, or a subroutine
Identify Operations
Static vs. Adaptive
Profile at Runtime
Simple Model
Simple vs. Complex
Predict Future Costs
Develop Power-Aware Execution Strategy
Our Four Models

• Lessons Learned
• Software counters can help modeling I/O behavior
• Capturing dynamicity is possible by coupling
  battery feedback and iterative linear regression
  models
• Conservative decaying of data is useful for
  improving model accuracy
• Model parameters should be chosen carefully to
  reduce multicollinearity
• Complex (many parameters) model estimations are
  more susceptible to feedback errors due to
  parameter dependencies

Extant Methods
OS Interfaces like ACPI Provides simple API to
battery voltage sensors Ok for different hw.
power levels - Not precise at all
HPMs Fast access Quite accurate -
Architecture dependent - Not designed for power
estimation --many events missing
Execution Time Simple to measure Fast and
precise - Not correlated to power - Not
suitable when hw. power levels change DVS, sleep