1. TinyOS: Concurrent Execution, and Power control
Class 9

2. Announcements Sign up for the weekend classes
Review phase 1 requirements
Each group update TA on your progress with TinyOS installation

3. Agenda TinyOS multi-threading Radio duty cycling
Power management in sensors

4. TinyOS multi-threading
TinyOS is a single application OS
It supports multi-threading for the application
Several threads can be posted in parallel which get scheduled in a FIFO queue
Tasks and events are primarily used for doing computation
Tasks are for long running computations
Events are used for very short computation and for posting long running computations in the FIFO queue

5. Concurrent execution uint16_t Samples[SAMPLES]; Uint16_t i = 0;
event void ReadStream.readDone(error_t ok, uint32_t SensedVal) {
if (ok == SUCCESS){
i++;
Samples[i] = Val;
if(i == SAMPLES-1){
i = 0;
post calculateAvg(); }
task void calculateAvg() {
uint16_t i, avg, var;
for (avg = 0, i = 0; i < SAMPLES; i++)
avg += Samples[i];
avg /= SAMPLES;
for (var = 0, i = 0; i < SAMPLES; i++) {
int16_t diff = Samples[i] - avg;
var += diff * diff;
In readDone, we need to compute the variance of the sample. We defer this “computationally-intensive” operation to a separate task, using post.
We then compute the variance

6. Exercise to understand concurrency
Lets have a loop in a task
Increment a counter
Have a timer fired event where at each fired event we send the counter value to the base station
What values do we see at the base station ?

7. Code Flow Diagram Boot Start Radio Call Timer
Booted
Start Radio
StartDone
Call Timer
Only the first task gets done
FIFO scheduler gives preference to events
TimerFired
Call Sensor
Task gets pushed back in the queue
ReadDone
Post 10 tasks

8. Radio Duty Cycling

9. How to efficiently duty cycle radio ?
Code from last class
Event based radio turn off
Not duty cycling
Duty cycling: Having a fixed schedule of radio off and on states
Depends on several factors:
Application communication and computation patterns
Emergency communication requirements
In case of networks, base station might sleep when a sensor sends a packet
Leads to packet loss
How to efficiently duty cycle radio ?

10. Lets go step by step Consider 3 sensors
One base station A
Two sensor nodes B and C
C communicates its data to B and B forwards it to A
The sensors periodically broadcasts messages to their neighbor
All the sensors want to sleep
Design a strategy to duty cycle the radio so that no packet is lost

11. Solution
Synchronous sleep
Sleep
Sensor C
Awake
Sensor B
Sensor A

12. Next Step Consider 4 sensors A collects data from B
Adjusts it sleep according to B’s sleep schedule
D collects data from C and adjusts its sleep according to C’s sleep schedule
Design a sleep schedule so that A can send data to D

13. Solution
For two sensors to communicate their the sender should know the receiver’s sleep schedule
Sensor B
Sensor A
Sensor D
Sensor C

14. Generic solution to duty cycling
Each sensor keeps a table of its sleep schedule
Each sensor then sends the table to all its immediate neighbors
The sensors then start their sleep schedule
When sensor A wants to communicate with B
A waits for B to wake up and then transmits
Once a transmission is started none of the sensors sleep until the transmission is finished
When finished both the sensors resume their sleep schedule

15. SMAC- Sensor Medium Access Control
Need an energy efficient MAC protocol
Design considerations –
Level 1 issues
Collision avoidance-a basic task of MAC protocols
Good scalability – sensors may die or new sensors maybe added
Energy efficiency
Often difficult recharge batteries or replace them
Prolonging the life-time is important
Level 2 issues
Latency, fairness, throughput, bandwidth

16. Collision Avoidance
When more than one sensor wants to send packets to a receiver there can be collision
Use two flags –
Request To Send (RTS) – If a sensor wants to send a packet to the receiver it sends out an RTS.
Clear To Send (CTS) – The receiver on reception of the first RTS sends out a CTS to the sender allowing access to the channel

17. Collision Avoidance Time Diagram
SYNC to synchronize sleep schedule
CS to sense the channel

18. Overhearing Avoidance
Problem: Receive packets destined to others
In , each node keeps listening to all transmissions from its neighbors for virtual carrier sensing
Each node should overhear a lot of packets not destined to itself
Solution: Letting interfering nodes go sleep after they hear an RTS or CTS packet
Which nodes should sleep?
All immediate neighbors of sender and receiver
S-MAC lets interfering nodes go to sleep after they hear an RTS or CTS
DATA packets are normally much longer than control packets
How long?
The duration field in each packet informs other nodes the sleep interval
After hearing the RTS/CTS packet destined to a node, all the other immediate neighbors of both the sender and receiver should sleep until the NAV becomes zero

19. Message Passing
Problem: Sensor net in-network processing requires entire message
Solution: Don’t interleave different messages
Long message is fragmented & sent in burst
RTS/CTS reserve medium for entire message
Fragment-level error recovery — ACK
— Extend Tx time and re-transmit immediately if no ACK is received
Advantages
Reduces latency of the message
Reduces control overhead
Disadvantage
Node-to-node fairness is reduced, as nodes with small packets to send has to wait till the message burst is transmitted

20. Aim to reduce energy Main reason for radio sleep is to save energy
Why save energy ?
What are the different energy consumers ?
How to save energy ?

21. Power Management

22. Sources of Power Dissipation
Sensing + Computation + Communication
Often more than computation
Least Power consuming
Most Power consuming
Power consumption depends on the type of processing
Sending <
Receiving
Communication energy more than 100 times greater than computation energy

23. Why power management ? Safety Sustainability
Energy dissipation causes temperature rise
Sensors on body can cause physical harm
Sensors (Mean Time To Failure) MTTF decreases
Sustainability
Scarce and intermittent energy sources
Often energy is scavenged from the environment
Energy efficiency to reduce power supply needs
Matching power needs

24. Power management strategies
Sensing
E.g. pulse oximeter operation causes temperature rise on human skin
Temperature rise depends on the sensing frequency
Reduce the frequency to control temperature
Compressed sensing

25. Power management strategies
Computation
Power consumption depends on different types of computation
Duty cycling of processor
Sleep states
Frequency scaling
Voltage scaling
Scheduling computation when energy is available
Predicting workload characteristics
Periodic vs asynchronous
Periodic workload enables efficient solutions
Async workload requires prediction which may fail often
Making energy available during peak workloads

26. Power management strategies
Communication
Sending and receiving requires power from the energy supply
Communication power > 100 X computation power
Receiving power more than sending
Low power listening
Radio duty cycling

27. Application dependency
Such strategies are however application dependent
Cannot turn off radio when the sensor needs to send
When an event occurs the radio has to wake up from low power state
Example Ayushman health monitoring system
SpO2
EKG
EEG BP
Base
Station
Motion
Sensor
Body Sensor Network

28. Ayushman workload
Example: Ayushman health monitoring application is considered as the workload
Sensing Phase – Sensing of physiological values (Plethysmogram signals) from the sensors and storing it in the local memory
Transmission Phase – Send the stored data to the base station in a single burst
Security Phase – Perform network wide key agreement for secure inter-sensor.
The Security phase occurs once in a day
The Sensing phase and Transmission phase alternate forming a sleep cycle
Sensor CPU Utilization
Time
Sensing Phase
Transmission Phase
Security Phase
Sleep Cycle
Ayushman Workload
Enables Sleep Scheduling
Frequency Throttling during security phase

29. Effects of Power Management
Periodic sensing and computation application in Ayushman
Sender Side: Sensing + FFT + peak detection + quantization + polynomial evaluation + random number (chaff) generation + data transmission + acknowledgement of transmission
Receiver Side: Sensing + FFT + peak detection + quantization + listen for packet + interpolation + transmit acknowledgement 1 2 3 4 5 6 7 8 9
0.01
0.02
0.03
0.04
0.05
0.06
Power Profile (Radio-ON)
Receiver(Radio ON)
Sender(Radio ON)
Sensing
FFT
Peak
+ Quant
Poly Gen +
Eval
Add Chaff
Vault Tx/Rx
Lagrangian
Interpolation
Ackn Tx/Rx 1 2 3 4 5 6 7 8 9
0.01
0.02
0.03
0.04
0.05
0.06
Power (mW)
Power Profile (Radio-OFF)
Receiver(Radio OFF)
Sender(Radio OFF)
Ackn Tx/Rx
Lagrangian
Interpolation
Vault Tx/Rx
Add Chaff
Poly Gen +
Eval
Peak
+ Quant
FFT
Sensing

30. Energy Savings Sensing power >> computation power
PKA Computation Energy Consumption for Different Vault Sizes
0.7
Sender(Radio ON)
Sender(Radio OFF)
0.6
Receiver(Radio ON)
Receiver(Radio OFF)
Sensing power >> computation power
Sensing
0.5
0.4
Energy (Joules)
~100 mJ savings when radio turned off
0.3
0.2
0.1
1000
2000
3000
4000
5000
Vault Size (Chaff Points)

31. Reduction of data transmission
To further reduce communication power we may avoid data transmission
Event based data transmission
Consider the context aware radio example in class
Light intensity below 10 is not interesting
So don’t send data
Another example –
Temperature data does not change too often
Quiz: What is an intelligent way to reduce transmission ?
Solution: Transmit only when there is a significant difference
Learn the characteristic of the sensed signal
Predict unusual changes
Send only when there are changes

32. Complex signals What happens when the signal is complex
Example: Electrocardiogram
Represents electrical activity of the heart
Consists of P, Q, R, S, T waves
The beat morphology and R-R intervals in an ECG are considered important for diagnosis

33. Generative Models Actual Signal
Need to learn the parameters for a given signal
Given the parameters the model should be able to regenerate the signal
Each wave
represented by
a Gaussian function
z =

34. Event Based data transmission
Normal
Expected Change
Unexpected Change
Morphology and inter beat features do not change
Only inter beat features change
Both morphology and inter beat features change
No communication
Use a generative model to represent ECG
Send inter beat feature updates
Send every sample

35. The Methodology Sensor matches sensed signal with model
If a match don’t send data
If model parameters vary, send only parameters
If unknown change occurs send sample by sample
At the base station use model to regenerate and align with raw signals to show the ECG

36. Energy and Memory Savings
Results
Implemented using off-the-shelf sensors
Base case for comparison: Send entire ECG
42:1
93%
Energy and Memory Savings
Diagnostic Accuracy

37. Demo Video

38. Lessons Learned Communication more expensive than computation
Communication duty cycling is an useful tool
Application dependent duty cycling
Event based communication enables lot of energy savings
Can only be applied to cases where the sensed signal has a definite pattern

39. Energy savings in computation
Frequency throttling
Voltage scaling
Example BSN with Atom processors
Allows Advanced Configuration and Power Interface (ACPI) control for frequency and voltage

40. Atom background Ultra low power processor for embedded applications
However, order of magnitude higher power dissipation than the state-of-art BSN node
IA-32 microarchitecture helps in easy application development
Can use high level programming languages to develop applications
Six low power sleep states with ultra low power deep sleep state
Sleep scheduling can be employed to reduce power consumption
Intel Speed Step technology enables seven different operating frequency levels
Clock frequency control to reduce operating power
Sleep state and frequency control performed through easy ACPI support (through Model Specific Register (MSR) accesses)
IA 32 architecture allows easily application development
Low power sleep states enable sleep scheduling
Intel speedStep enables frequency scaling

41. Strategies to Address – Safety and Sustainability Challenges
Challenge - Atom’s high TDP (2.2 W) with respect to present day sensor nodes (~ 80 mW [4])
Remedy – Power budgeting through sleep scheduling and clock frequency control
Road Blocks –
In a sleep mode the processor cannot compute
Decrease in clock frequency increases computation time
Challenge - Increase lifetime of operation
Remedy – include scavenging nodes in the BSN that will charge the Atom nodes wirelessly and supplement battery
The operation of scavenging sources are intermittent depending on the stochastic behavior of the host
Often the scavenging nodes fail to provide appropriate power levels to the nodes
Stress on the point that the power budgeting schemes are strongly dependent on the application real time requirements
The strategies are closely related to the applications real time requirements.
Intelligent design is required to achieve safety and sustainability
while respecting the real time requirements of the applications
4. K. Venkatasubramanian et al. Green and sustainable cyber-physical
security solutions for body area networks. In BSN ’09: Proc. of the Sixth
Intl. Workshop on Wearable and Implantable Body Sensor Networks,
pages 240–245, Washington, DC, USA

42. BSN Hardware model BSN node Base Station Scavenging Sources BSN Node
Intel N270 single core processor
1.6 GHz clock frequency, 1 GB RAM
Intel SpeedStep frequency control technology – useful for power management
6 sleep states including one ultra low power sleep state (C6) – sleep scheduling
Chipcon 2420 radio
2.45 GHz, wireless standard
Maximum Power dissipation (58 mW [4])
BSN Node
Base Station
Base Station
Atom based mobile phone
Wireless Charging
Scavenging Sources
Body Heat, Ambulation, Respiration and Sun Light
Wireless charging of BSN nodes from scavenging sources is assumed
Each source has a specified range upto which it can charge nodes
Scavenging Sources

43. BSN node Software
The power consumption of Atom processor depends on the Operating System used
Mobile Intel 945 GMCH board power consumption
Open Suse Linux = 11.7 W
Moblin OS = 10.4 W
ACPI support required for accessing Intel SpeedStep frequency control and sleep states
Moblin provides ACPI through which one can write to or read appropriate MSR registers to –
Control clock frequency
Sleep States
Measure core temperature
The BSN workload considered is the Ayushman application

44. Model based Analysis Model: An abstraction in a mathematical domain
Behavioral models (e.g. control system transfer function)
Consider the system as a black box
Find the relation between the inputs and the outputs
Formal models (e.g. finite state machines, petrinets)
Represent the operation of the system in terms of states and transitions between them
Structural models (XML specifications)
Represent the structure of a system in terms of the basic blocks that are present in the system
Does not necessarily represent operational characteristics

45. Model based analysis
Model based analysis normally used to verify critical systems such as avionics.
no need for actual scenario generation putting lives/property at risk.
Formal models for abstraction of the system behavior.
Expected system properties depend on the requirements.
Formal models analyzed through model checking to verify the system properties.
We use model based analysis to evaluate effectiveness of crisis response processes.
System Profiling
System Requirements
Models
Expected Properties
Model
Checking
Property Verification
Requirement Verification

46. Profiling Requirements
Thermal Safety – The maximum temperature of the skin in contact with the node should not exceed 39 ºC for 24 Hrs of operation
Thermal behavior of Atom under the given workload has to be evaluated
Sustainability – The available power from the scavenging sources should be able to meet the power demands of Atom node under the given workload
Power profiling of Atom processors during execution of Ayushman

47. Thermal Profiling
Requires core temperature measurements for different operating points of the Atom processor
The Mobile Intel 945 GSE development platform (GMCH) provided by Intel has digital thermal sensors
The board thermal sensors were read from Model Specific Registers
The maximum core temperature (43 ºC) was observed during PKA execution
Turn On GMCH board
Set Operating
Frequency
Read MSR
C6 Sleep State
Run Ayushman
Log Temperature Data
Thermal profiling methodology

48. Power Profiling
PKA is the most power consuming computation in Ayushman [3]
The difference between idle power and power during PKA execution was measured using the GMCH board
Idle power of Atom N270 processor was added to it to obtain PKA power consumption
AC Mains
Power Meter
Intel Atom N270 on Mobile Intel Chipset 945 GSE
Board Power Lead
Operating Mode (Percent throttling)
Power (W)
0.191 13
0.1864 25
0.17
37, 50, 62, 75
0.167 87
0.164
Power Measurement Set up
Table showing Atom power consumption for PKA execution at different operating frequencies

49. Resource Consumption
PKA computation in Ayushman involves signal processing of physiological signals as well as execution of security algorithms
Resource footprint of PKA is evaluated in terms of –
RAM usage, Power Consumption, and Computation Time
Atom compared to TelosB provides very low RAM usage and computation time
However as expected it has around thrice the power consumption
Platform
RAM Usage
Power (mW)
Computation Time (ms)
TelosB
80% 66
2186
Atom
0.006%
164 41
Resource Consumption for PKA execution

50. Safety Analysis
The temperature rise of human skin due to contact with Atom based BSN node has to be evaluated
The temperature rise occurs due to several physical phenomenon and is modeled using the Penne’s bioheat equation -
Skin Temperature
Radio Power
Atom Power Consumption
Atom Operating Temperature
Heat
accumulated
Heat transfer
by conduction
Heat transfer
by convection
Heat by electromagnetic
radiation
Heat by power dissipation
Heat by
radiation
Thermal damage parameter calculated according to Henrique and Moritz [5]. Maximum temperature must not exceed this.
Operating Mode (Percent Throttling)
Atom Processor Operating Temperature (ºC)
Maximum Skin Temperature (ºC) after 24 hrs of operation
Thermal Damage Temperature (ºC)
0, 13, 25 43
39.2
37, 50 42
62, 75 41 87 39
5. F. C. J. Henriques et al. Studies of thermal injury: I. the conduction of heat to and through skin and the temperatures attained therein. A theoretical and an experimental investigation. In Am J Pathol., pages 530–549, July

51. Sustainability Analysis
Duty Cycling of Atom operation during Ayushman execution
Sleep mode (C6) during Sensing Phase
Power consumption = Psleep for time ts
Active mode during data transmission phase
Power consumption = Pactive + radio power Pradio for data transmission time ttx
Active mode during PKA execution
Power Consumption = Pactive + PKA execution power PPKA for time tPKA
PKA involves transmission of security related information (vault) between two sensors
The Atom processor must be in active state with the radio on. Power Consumption = Pactive + Pradio for time tvault
Total energy consumption for n BSN nodes
x is the number of sleep cycles required in a day
Total Energy
Sensing Energy
Transmission Energy
PKA Computation Energy (Pair wise PKA)
PKA communication Energy (Pair wise PKA)
Total Sleep Cycle Time
Pair wise PKA Execution Time
Sensor CPU Utilization
Time
Sensing Phase
Transmission Phase
Security Phase
Sleep Cycle
Ayushman Workload
Enables Sleep Scheduling
Frequency Throttling during security phase

52. Sustainability Analysis Results
Four energy scavenging sources were considered
Body Heat, Ambulation, Respiration and Sun Light
The total energy obtained from any combination of scavenging sources should provide EBSN amount of energy for n nodes
Scavenging Source
Available Power (W)
Scavenge Time (Hrs)
Body Heat
0.1 – 0.15 24
Ambulation
1.5 2
Respiration
0.42 6
Sun Light
0.1 3
Combination of Scavenging Sources
No of Sustained BSN nodes
Body Heat and Respiration 25
Ambulation and Respiration 40
All four
110

53. Thank You