dynamic energy burst scaling for transiently powered systems · a gomez, et al. dynamic energy...
TRANSCRIPT
Andres Gomez1, Lukas Sigrist1, Michele Magno1,2, Luca Benini1,2, Lothar Thiele1
1D-ITET, ETH Zurich, Switzerland 2DEI, University of Bologna, Italy
Dynamic Energy Burst Scalingfor Transiently Powered Systems
A Gomez, et al. Dynamic Energy Burst Scaling
2
15.03.2016
http://www.electronicsweekly.com/new
s/research-news/device-rd/isscc-rice-
grain-solar-sensor-nodes-uses-cortex-
m3-2010-02/
http://www.seeedstudio.com/depot/Wireless-Sensor-Node-Solar-Kit-p-919.html
http://www.prnewswire.com/news-releases/cypress-
introduces-the-worlds-lowest-power-energy-harvesting-
power-management-ics-for-battery-free-wireless-sensor-
nodes-300130529.html
http://senseonics.com/product/the-sensor
Design Trends
Battery-based Batteryless
A Gomez, et al. Dynamic Energy Burst Scaling
3
15.03.2016
What are Transiently Powered Systems?
Energy-harvesting based systems with limited energy storage capacity
Why harvesting based?
• Low-cost
• Long-term
• Environmentally friendly
Why limit energy storage?
• Batteries (and supercaps) self-discharge, limited cycles
• Storage is expensive in terms of cost, form factor
• Storage requires energy surplus
A Gomez, et al. Dynamic Energy Burst Scaling
4
15.03.2016
Energy-driven embedded systems that:
• Have low cost
• Operate efficiently
• Guarantee program progress
Challenging Scenario:
• Low power harvesting conditions
Transient System Design Goals
eventTransient
Node
harvest energy
sense information
Transient Camera
actuate
Light Source
A Gomez, et al. Dynamic Energy Burst Scaling
5
15.03.2016
Directly Coupled Systems
Load
[Balsamo2015]
Advantage:
• Simplicity, no additional sources of leakage/losses
Disadvantage:
• Source must have the same operating power point as the load
Federating Energy:
[Hester2015]
“Checkpointing” for MCU’s:
A Gomez, et al. Dynamic Energy Burst Scaling
6
15.03.2016
Boost-Buck Converter Topology
Advantages:
• Source at Maximum Power Transfer Point
• Load has regulated voltage
Disadvantage:
• Introduces new sources of losses (converter inefficiencies, leakage, etc)
• No tracking of load’s optimal operating point
• Decoupled voltages:
𝑃𝑙𝑜𝑎𝑑
𝑉𝑙𝑜𝑎𝑑Load
𝑃𝑖𝑛
𝑉𝑖𝑛Source
BoostConverter
BuckConverter
[Naderiparizi2015]
StorageSource Load
Energy
Burst
WispCam:
A Gomez, et al. Dynamic Energy Burst Scaling
7
15.03.2016
Our Proposed System
Energy Management Unit (EMU):
• Based on the Boost-Buck topology
• Optimized storage element
• Minimized wake-up time, cold-start energy
• Tracks load’s optimal operating point
• Feedback-based Dynamic Energy Burst Scaling
A Gomez, et al. Dynamic Energy Burst Scaling
8
15.03.2016
How can we design transient applications?
Basic Sense (S) and Process (P) Application
• Consists of two atomic tasks: S and P
• Known energies (ES, EP) and operating voltage (Vload)
• Two burst-based ways of executing this application: together, individually
Single Burst @ Vload
Estored
t
buffer ES + EP
Multiple Bursts @ Vload
Estored
t
buffer ES buffer EP
• Minimized buffer sizes
• Minimum start-up time
• Task-level optimization
• Minimized buffer sizes
• Minimum start-up time
• Task-level optimization
• Higher buffer sizes
• Longer timing constants
• Application-level optimization
• Higher buffer sizes
• Longer timing constants
• Application-level optimization
A Gomez, et al. Dynamic Energy Burst Scaling
9
15.03.2016
Minimizing the Storage Element
Usable energy stored
between [𝑉𝑙𝑜𝑎𝑑 , 𝑉𝑚𝑎𝑥]
• For a single-capacitor system, 𝐶𝑚𝑖𝑛 ensures functionality
𝐶𝑚𝑖𝑛 =2 max(𝐸𝑡𝑎𝑠𝑘,𝑖)
𝜂𝑏𝑢𝑐𝑘 ∗ (𝑉𝑚𝑎𝑥2 − 𝑉𝑙𝑜𝑎𝑑
2 )
• 𝐶𝑚𝑖𝑛 also miminimizes cold-start energy and start-up time
𝑡𝑠𝑡𝑎𝑟𝑡−𝑢𝑝 = 𝑡 | 𝑉𝑐𝑎𝑝 𝑡 =2 0
𝑡𝐸𝑐𝑎𝑝′ 𝜏 𝑑𝜏
𝐶𝑐𝑎𝑝= 𝑉𝑙𝑜𝑎𝑑
A Gomez, et al. Dynamic Energy Burst Scaling
10
15.03.2016
Energy Burst Scaling
Basic Sense (S) and Process (P) Application:
• S: minimum voltage VS=3V (peripheral req.)
• P: minimum voltage VP=2V
Constant Bursts (WispCam): Dynamic Bursts (Proposed):
Estored
t
Vload=max(VS,VP)
S+P
Estored
t
Vload=VS Vload=VP
S P
To minimize load energy:
𝐸 = 𝑽𝒍𝒐𝒂𝒅 ∗ 𝐼𝑙𝑜𝑎𝑑 ∗ 𝑡 ⇒ 𝐸𝑚𝑖𝑛 = 𝑽𝒍𝒐𝒂𝒅,𝒎𝒊𝒏 ∗ 𝐼𝑙𝑜𝑎𝑑 ∗ 𝑡Dynamic Energy Burst Scaling:
+ Minimizes task energy
- Requires control mechanism!
Dynamic Energy Burst Scaling:
+ Minimizes task energy
- Requires control mechanism!
A Gomez, et al. Dynamic Energy Burst Scaling
11
15.03.2016
Dynamic Energy Burst Scaling
Feedback Loop:
• Minimizes task energy required to execute application
• Load configures EMU to provide Eburst @ Vload
• Task execution is triggered by EMU interrupt
A Gomez, et al. Dynamic Energy Burst Scaling
12
15.03.2016
Energy Burst-Based Flow
𝑃𝑙𝑜𝑎𝑑
𝑡
𝑉𝑐𝑎𝑝
𝑡
deep sleepactiv
e
activ
e
VSVP
deep sleep
EMU
interrupt
EMU
interrupt
Pload=60 nW Pload= ~4 mW
Estored (t) = ES Estored (t) = EP
Pload=60 nW Pload= ~3 mW
A Gomez, et al. Dynamic Energy Burst Scaling
13
15.03.2016
Modeling Transient Applications
task set
capacitance
EMU
Modelpower consumption
Pin(t)Vcap(t)
Eload / Eharvested
Pload(t)
Discrete-Time Simulation:
• Calculate voltage changes in capacitor (with non-ideal converters, leakage, etc)
• Verify understanding of EMU behavior
Inputs:
• Source’s power trace
• Etask, Vtask, Ptask
Output:
• Energy Efficiency
• Total Executions
A Gomez, et al. Dynamic Energy Burst Scaling
14
15.03.2016
Experimental Set-Up
Energy Management Unit
Control Circuit
HarvesterBQ25505
BuckTPS62740
Optimal Cap80 μF
ControlInterface
𝑷𝒊𝒏 𝑷𝒍𝒐𝒂𝒅
𝑽𝒍𝒐𝒂𝒅Load
MP3-25Solar Panel 𝑽𝒊𝒏
Load:
• Sensor: Centeye Stonyman (3V)
• Processor: MSP430 (2V)
Metrics:
• Energy harvested
• Energy consumed by the load
• Application execution rate
Control Signals
Energy Flow
PEMU=~10 uW
A Gomez, et al. Dynamic Energy Burst Scaling
15
15.03.2016
Constant Bursts - Characterization
Stage Voltage [V] Time [ms] Energy [μJ]
Boot 3.0 3.20 6.4
Sense 3.0 47.0 174.0
Process 3.0 50.6 220.8
Total 100.8 401.2
Estored
t
Vload=3V
S+P
ES+P=401.2 μJ
Pavg=3.98 mW
A Gomez, et al. Dynamic Energy Burst Scaling
16
15.03.2016
Dynamic Bursts - Characterization
1st Burst = Sense
Phase Voltage [V] Energy [μJ]
Boot 3.0 6.1
Sense 3.0 183.9
Total 190.0
2nd Burst = Process
Phase Voltage [V] Energy [μJ]
Boot 2.0 3.1
Process 2.0 143.7
Total 146.8
Constant Bursts: 401.2 μJ required per execution
Dynamic Bursts: 336.8 μJ required per execution
Constant Bursts: 401.2 μJ required per execution
Dynamic Bursts: 336.8 μJ required per execution
Estored
t
Vload=3V Vload=2V
S P
ES+P=(190 + 146.8) μJ = 336.8 μJ
Pavg,S+P= ~3.34 mW
A Gomez, et al. Dynamic Energy Burst Scaling
17
15.03.2016
Experimental Evaluation
EMU Testing Methodology:
• Constant Bursts Experiment
• Dynamic Bursts Experiment
• Run simulations with recorded input power traces
• Compare measured/simulated outputs
Harvesting Scenarios:
• Constant/variable input power
• Low power ranges: 0 – 400 μW
A Gomez, et al. Dynamic Energy Burst Scaling
18
15.03.2016
EMU System Efficiency (Eload/Eharvested)
Efficiency is bounded by converter efficiencies (~78 %)Efficiency is bounded by converter efficiencies (~78 %)
Dynamic Bursts Constant Bursts
A Gomez, et al. Dynamic Energy Burst Scaling
19
15.03.2016
Application Execution Rate
Dynamic Bursts: ~50 exec/min @ 400 uW
Constant Bursts: ~40 exec/min @ 400 uW
Dynamic Bursts: ~50 exec/min @ 400 uW
Constant Bursts: ~40 exec/min @ 400 uW
Dynamic Bursts Constant Bursts
A Gomez, et al. Dynamic Energy Burst Scaling
20
15.03.2016
Variable Input Power Experiment
Constant Bursts
Avg. Pin Metric Simulation Experiment
111.9 μWExec. Rate 9.87 min-1 9.93 min-1
ηsystem 67.76 % 68.01 %
Dynamic BurstsAvg. Pin Metric Simulation Experiment
92.3 μWExec. Rate 9.93 min-1 10.33 min-1
ηsystem 66.11 % 68.82 %
Experiment 1:
Experiment 2:
Camera worn around the lab for 15 mins. in various lighting conditions
A Gomez, et al. Dynamic Energy Burst Scaling
21
15.03.2016
Summary
Proposed Energy Management Unit (EMU):
• Large operating ranges (source/load decoupling)
• Minimized storage element, wake-up time, cold-start losses
• Optimized Power Point Tracking for source and load
• Feedback-based technique (Dynamic Energy Burst Scaling)