Title: A Unified Power Management Framework for Wireless Sensor Networks
1A Unified Power Management Framework for Wireless
Sensor Networks
Guoliang Xing Department of Computer Science and
Engineering Washington University in St.
Louis http//www.cse.wustl.edu/xing/
2Wireless Sensor Network Platform
- Integration of sensing, computation, and
communication - Example Mica2 mote
- Radio lt 40 Kbps
- Memory 4KB data, 128 KB program
- Limited power source 2AA batteries
- Several days of lifetime if continuously active
3Wireless Sensor Network Applications
Healthcare
Structural monitoring
Habitat monitoring
Perimeter security
- Limited power supplies batteries, small solar
panels - Long lifetime requirement months to tens of
years - Must minimize the total network energy consumption
4A Unified Power Management Framework
- An analytical model for minimizing total energy
consumption in all radio states - A new power management protocol that integrates
sleep scheduling with power control - A system architecture for flexible power
management
5Understanding Radio Power Cost
Radio power consumption in different states
(unit mW)
- Sleeping consumes much less power than idle
- Reduce idle energy of non-communicating nodes
through sleeping - Motivate sleep scheduling Polastre et al. 04, Ye
et al. 04 - Transmission consumes most power
- Reduce transmission energy of communicating nodes
- Motivate transmission power control Singh et al.
98,Li et al. 01,Li and Hou 03 - None of existing schemes minimizes the total
energy consumption in all radio states
6An Example of Minimizing Total Radio Energy
c
- a sends to c at normalized rate of r Date Rate
/ BandWidth - Source and relay nodes remain active
- Configuration 1 a ?c, b sleeps
-
- Configuration 2 a ? b ? c
b
a
node as avg. power
node bs avg. power
node cs avg. power
7Key Observations
Transmission power dominates short radio range
is preferable
Idle power dominates long radio range is
preferable since more nodes can go to sleep
8Problem Formulation
- a communicates with c at rate r, b sleeps
ignore sleeping power
group rate related terms
Ca,cPtx(a,c)Prx-2Pidle
edge (a,c) has a cost of Ca,c per unit of data
each active node has a cost of Pidle
Pidle
Ca,c
- Extending to the general case.
c
Pidle
a
b
rate r
9Minimum Power Configuration (MPC)
- Given traffic demands I( si , ti , ri ) and
G(V,E), find a sub-graph G(V, E) minimizing
independent of data rate!
sum of edge cost from si to ti in G
- Sleep scheduling
- Power control
- Sleep scheduling
- Power control
10Solutions
- Minimum Power Configuration is NP-hard
- Matching Based Algorithm (MBA) can solve the
one-sink case with approx. ratio O(lgk) Meyerson
et al. 00 - Distributed implementation is expensive
- New Incremental Shortest-path Tree Heuristic
(ISTH) - Distributed and online
- Known approx. ratio is O(k), similar average-case
performance to MBA
11Incremental Shortest-path Tree Heuristic (ISTH)
- Initially, all nodes are labeled as asleep
- For each traffic demand (si, ti,ri)
- Find the shortest path from si to ti under edge
cost H(u,v, ri) - Label all nodes on the found path as active
- When a node is asleep, H includes both node cost
Pidle and edge cost riCu,v - When a node is active (on an existing path), H
encourages path reuse by removing node cost Pidle
12Illustration of ISTH
- Edge cost Cu,v2, node cost Pidle1
- Find a new path in each iteration
1
2
cost reduction!
2
1
1
1
0
2
2
New cost
2
1
1
1 2 0.2 2 22.8
2
1
2
2
Source 2 r2 0.2
1
1 0.2 2 21.8
0
1
Source 1 r1 0.2
1 2 0.2 2 22.8
13Minimum Power Configuration Protocol (MPCP)
- Extends the DSDV protocol with the routing metric
H(u,v,ri) - Routing cost is dependent on data rate and state
of the node (asleep or active) - Uses different transmission power to different
neighbors - Activates a node if on a route, schedules to low
duty cycle otherwise
14Simulation Environment
- Prowler simulator extended by Rmase project
- Prowler http//www.isis.vanderbilt.edu/projects/n
est/prowler/ - Rmase http//www2.parc.com/spl/projects/era/nest/
Rmase/ - Implemented USC model Zuniga et al. 04 to
simulate lossy links of Mica2 motes - Baseline protocols
- MT Extended DSDV that minimizes num of Txs
- MTP Extended DSDV that minimizes Tx Power
- Data rate per flow 0.3 Kbps, 100 nodes
15Energy Cost
Energy Cost of All Nodes (J)
Energy Cost of Non-Source Nodes (J)
Energy cost of all nodes MPCP saves as much as
30 energy
Energy cost of non-source nodes MPCP saves as
much as 80 energy
16A Unified Power Management Framework
- An analytical model for minimizing total energy
consumption in all radio states - A new power management protocol that integrates
sleep scheduling with power control - A system architecture for flexible power
management
17Architectural Issues
- Sleep scheduling is coupled with the MAC layer
- Difficult to implement new sleep schedulers
- Sleep scheduling evolution dependent on MAC
- Potential conflicts from different upper-layer
protocols - How to coordinate two power management protocols?
Protocol 0
Protocol 1
Protocol 2
Protocol 3
Protocol 4
S-MAC
802.15.4
B-MAC
Sync Sleep Scheduling
TDMA Scheduling
Low Power Listening
PHY
Sleep Scheduling in TinyOS
18Architectural Requirements of Power Management
- Flexibility
- Power management protocols should be independent
of other system layers (e.g., app, routing, MAC) - User can specify desirable power management
policies - Composibility
- Combine different power management protocols for
different applications - Cross-layer optimization
19Unified Power Management Architecture (UPMA)
- Power management abstraction
- Allow each user to specify desirable power
management polices - Power manager
- Aggregate multiple users parameters to a same
policy - Coordinate the use of multiple policies
- Interfaces between power management and other
layers (e.g., MAC, routing)
20UPMA -- Sleep Scheduling
interfaces of sleep schedulers
Protocol 2
Protocol 1
Protocol 3
Protocol 0
RadioDutyCycling
LowPowerListening
Other Interface
parameters specified by upper-level protocols
OnTime
Mode
Param 0
OffTime
Preamble
Param 1
DutyCycling Table
LPL Table
Other Table
Power Management Abstraction
aggregate parameters in the tables
Power Manager
Aggregator
sleep scheduling protocols
Low Power Listening
Others
Basic Sleep Scheduler
MAC
PreambleLength
ChannelMonitor
On/Off
interfaces with MAC
PHY
21Case Study Duty Cycle Backbone
- Duty cycling application
- turn on radio for 2s to report current
temperature in every 10s - PEAS Ye et al. 03
- One active node in any 10-meter range, other
nodes run in power-saving mode (turn on radio for
1s in every 25s) - After aggregation
- In any 10-meter range, only one node reports
temperature every 10s, other nodes turn on radio
for 1s in every 25s
application duty cycle
PEAS low duty cycle
PEAS low duty cycle
combined duty cycle
22Instantiation of UPMA
Duty Cycling App
PEAS
RadioDutyCycling
Aggregator if OnTime 8, combine (2,8) and
(1,24) else, run (1,24)
Power-saving / Active
Duty Cycling
OnTime
1 / 8
2
OffTime
8
24 / 0
Power Management Abstraction
Power Manager
Basic Sleep Scheduler
MAC
ChannelMonitor
On/Off
PHY
23Implementation
- Implemented UPMA in TinyOS 2.0 for both Mica2 and
Telosb motes - Developed interfaces with different types of MAC
- CSMA based S-MAC Ye et al. 04, B-MAC Polastre
et al. 04 - TDMA based TRAMA Rajendran et al. 05
- Hybrid 802.15.4, Z-MAC Rhee et al. 05
- Separated sleep scheduling modules from B-MAC
- Implemented two new sleep schedulers on top of
B-MAC
24Evaluation
PEAS only
- 15 Telosb motes run PEAS, and one of the 6 duty
cycles - Active nodes send a packet to base station during
the on time - Instrumented the radio stack of CC2420 to account
the total time the radio stays in each state
25Research Summary
- An unified power management framework TOSN 06,
MobiHoc 05, tech. report 06 - Spatiotemporal query service for mobile users in
mostly sleeping sensor networks ICDCS 05, IPSN
05 - Integrated power management under both sensing
and communication constraints - Fundamental relation between connectivity and
sensing coverage TOSN 1(1), SenSys 03 - Impact of sensing coverage on routing performance
TPDS 17(4), MobiHoc 04 - Detection-based data fusion IPSN 04
- nORB Light-weight real-time middleware for
networked embedded systems RTAS 04
26Conclusions
- A unified power management framework
- MPCP first protocol that jointly optimizes the
total energy consumed in all radio states - UPMA first unified power management architecture
for wireless sensor networks - Laid foundation for flexible and composable power
management
27Future Work
- Integrated power management
- Radio, CPU, flash, sensors
- Tools for configurable power management
- Power management in heterogeneous networks
28Acknowledgements
- Washington University in St. Louis
- Advisor Chenyang Lu
- Collaborators Robert Pless, Gruia-Catalin Roman,
Sangeeta Bhattacharya, Octav Chipara, Chien-Liang
Fok, Kevin Klues - Palo Alto Research Center (PARC)
- Ying Zhang, Qingfeng Huang
29Overhead
- Most MPCP route updates are local ? overhead
remains roughly constant as num of flows grows
30Minimum Steiner Tree Heuristic (MSTH)
- The algorithm
- Assign the cost of each edge to be Pidle
- Run a distributed minimum Steiner tree
approximate algorithm - Properties
- The approximate ratio is about
( 5 on mica2 motes) - Good performance when Pidle is high
31Greedy Prefetching
Uninvolved nodes
Collector node
Active nodes
Alerted nodes
- Forward a prefetch msg ASAP
- Many query areas are set up simultaneously
- High network contention storage cost
- Prediction to pickup points far away is likely
wrong
32Just-in-time (JIT) Prefetching
Uninvolved nodes
Collector node
Active nodes
Alerted nodes
- Forward a prefetch msg at the right time
- Only a few query areas are set up simultaneously
- Reduced network contention storage cost
- More robust to user motion changes
- Implemented on Mica2 motes, demoed at SenSys 04
33Performance of Centralized Algorithms
- TMST Min spanning tree Li et al. 2003
- TSPT Shortest path tree singh et al. 1998
- MBA the Matching Based Algorithm Meyerson et
al. 00 - MBA-opt MBA with our optimization
- 200 nodes distributed in a 500m X 500m region
- Each data flow has a rate of 0.2 Kbits/s
- Radio uses Mica2 mote setting
34Evaluation IIAggregation Performance
Up to 6 duty cycles On time200ms Off
time0.2s, 0.6s, 1.4s, 3s, 6s, 12.6s Each slave
node runs one of the duty cycles, the master node
runs the aggregate duty cycle
Up to 6 duty cycles On time200ms Off
time0.2s, 0.6s, 1.4s, 3s, 6s, 12.6s Each slave
node runs one of the duty cycles, the master node
runs the aggregate duty cycle
35Properties of ISTH
- Distributed implementation is easy
- Known approx. ratio is k, num of sources
- Performance for special cases
- Approx. ratio is 2 when r 0
- Suggest good performance with low data rates
- Optimal when Pidle0
36Delivery Rate and Delay
- MPCP/MASP cause slightly higher network
contention due to more path reuse
37Use Case I Duty Cycle Aggregation
- Applications specify different duty cycles
through the abstraction component - Power manager aggregates different duty cycles
- Power manager runs the Basic Sleep Scheduler to
turn on/off radios
Duty Cycle 0
Duty Cycle 1
Aggregate Duty Cycle
38nORB Critical Path
39Critical Path Issue Discovered
GIOP Header
A
A
B
A
B
A
GIOP Payload
A
A
A
A
B
B
A
A
A
40Critical Path Issue Discovered Contd.
41nORB Critical Path Optimizations
- Client ORB
- Use gather-write pattern to send header payload
at the same time - Server ORB
- Issue one non-blocking read with large buffer
size(2K) - Check the integrity of request after read returns
- If the request is complete, dispatch it to the
servant. - If the request is not complete, return to the
reactor - If gets multiple requests, split and queue them
for later dispatching
42nORB Critical Path Optimization - Effect
Average latency reduction is 50
43Existing Approaches of Radio Power Management
- Sleep scheduling
- Duty cycling
- Nodes run in duty cycles (interleaving active and
sleeping intervals) - Backbone maintenance
- Keep a subset of nodes active and schedule others
to sleep - Transmission power control
- Topology control reduce per node power
- Power aware routing reduce per packet power
- Sleep scheduling in TinyOS
- MAC protocols S-MAC Ye et al. 04, B-MAC
Polastre et al. 04
44Problems with Existing Solutions
- Only reduce partial radio energy
- Sleep scheduling only reduces idle energy
- Power control only reduces TX energy
- Lack of flexibility and composibility
- Tight coupling with other system functions
- Sleep scheduling is often part of MAC
- Power control is often part of routing
- Difficult to replace existing or implement new
power management protocols - Different protocols cannot work together
efficiently
45Greedy Geographic Routing in Sensing-covered
Networks
- Greedy forwarding
- Chooses as the next hop the neighbor closest to
destination - Always succeeds in sensing-covered networks
- Bounded Voronoi Greedy Forwarding (BVGF)
- Combines greedy forwarding with Voronoi diagram
- Always finds routing paths with bounded lengths
destination
Rc
B
A
Closest to destination
46Topology Control for Wireless Sensor Networks
- Wireless links are inherently lossy
- Excessive packet loss and energy waste
- 50-80 transmission energy was wasted by packet
retransmissions zhao et al. 03 - When network workload is low
- Link failures are mostly due to path fading and
environmental noise - Higher transmission power leads to higher link
quality - Existing topology control algorithms do not
account for lossy links - Achieve required topology quality using minimum
transmission power