Accurate Prediction of Power Consumption in Sensor Networks
University of Tubingen, GermanyIn EmNetS 2005
Presented by Han
Outline
• Goal
• Approach to build AEON
• Power evaluation of TinyOS
• Comparison with PowerTossim
Goal
• To evaluate energy consumption of real codes– Algorithms and programming styles influence
power consumption– Predict network lifetime
Approach
• Build an energy model
• Implement the energy model in an emulator
• Use the emulator to analyze power consumption of real codes and verify
Building energy model
• Based on Mica2 platform
• Write special TinyOS programs to turn on each hardware component each time
• Measure the current draw
Energy model
Approach
• Build an energy model
• Implement the energy model in an emulator
• Use the emulator to analyze power consumption of real codes and verify
Implementation
• AEON is implemented on top of AVRORA
AVRORA
• Developed by UCLA (IPSN’05)
• Instruction-level simulator– Runs actual microcontroller program
• Tossim use software to model hardware components– Lose timing and interrupt properties
• AVRORA is 50% slower than Tossim
Approach
• Build an energy model
• Implement the energy model in an emulator
• Use the emulator to analyze power consumption of real codes and verify
Validation
• Average error 0.4%• deviation 0.24• Predict 172 hours for CntToLedsAndRfm
• 168 hours by Crossbow lifetime test
Blink application
Evaluation of Apps
Executed for 60 seconds
CntToLedsAndRfm
Radio interrupt (radio is not turned off between transmission)
Radio transmission
HPLPowerManagement
• Dynamically switch the CPU between six sleep modes based on the current load
Low power listening (B-MAC)
High data rate (wake up more frequently)
Low data rate (wake up less frequently)
Predicted savings
Energy profiling
• Map source code functions to the corresponding object code addresses (Surge)
PowerTossim
• Developed by Harvard (SenSys’04)
• Build on top of Tossim
• Based on nearly the same measurement
• Benefit from the scalability of Tossim
• Also lose some accuracy on capturing interrupts
Comparison
• For the same CntToLedsAndRfm application
• PowerTossim predicts 2620mJ/min
• AEON predicts 3023mJ/min
• AEON claims that the additional energy is spent on reloading counter after timer interrupt
Results from PowerTossim
Conclusion
• More accurate than PowerTossim (?)
• The energy evaluation parts give quantitatively improvement of designed protocols
• This tool would be useful in software development