Deborah Estrin

1
Embedded Networked Sensing Systems motivations
and challenges(Stanford University EE3925 1/6/04)

• Deborah Estrin
• http//cens.ucla.edu/Estrin
• destrin_at_cs.ucla.edu
• Work summarized here is largely that of students,
  staff, and other faculty at CENS
• We gratefully acknowledge the support of our
  sponsors, including the National Science
  Foundation, Intel Corporation, DARPA, Inc.,
  Crossbow Inc., and the participating campuses.

2
Why Embedded Networked Sensing ?

• Micro-sensors, on-board processing, wireless
  interfaces feasible at very small scale--can
  monitor phenomena up close
• Enables spatially and temporally dense
  environmental monitoring
• Embedded Networked Sensing will reveal previously
  unobservable phenomena

Contaminant Transport
Ecosystems, Biocomplexity
Marine Microorganisms
Seismic Structure Response
3
ENS enabled by Networked Sensor Node Developments
LWIM III UCLA, 1996 Geophone, RFM radio, PIC,
star network
AWAIRS I UCLA/RSC 1998 Geophone, DS/SS Radio,
strongARM, Multi-hop networks
Sensor Mote UCB, 2000 RFM radio, Atmel
Medusa, MK-2 UCLA NESL 2002

• Predecessors in
• DARPA Packet Radio program
• USC-ISI Distributed Sensor Network Project (DSN)

4
ENS Technology Design Themes

• Long-lived systems that can be untethered
  (wireless) and unattended
• Communication will be the persistent primary
  consumer of scarce energy resources (MICA Mote
  720nJ/bit xmit, 4nJ/op)
• Autonomy requires robust, adaptive,
  self-configuring systems
• Leverage data processing inside the network
• Exploit computation near data to reduce
  communication, achieve scalability
• Collaborative signal processing
• Achieve desired global behavior with localized
  algorithms (distributed control)
• The network is the sensor (MangesSmith,
  Oakridge Natl Labs, 10/98)
• Requires robust distributed systems of hundreds
  of physically-embedded, unattended, and often
  untethered, devices.

5
ENS Architecture Drivers
DRIVERS
TECHNICAL CAPABILITIES
Adaptive Self-Configuring Wireless Systems
Varied and variableenvironments
Energy and scalability
Distributed Signal and Information Processing
Heterogeneity of devices
Networked Info-Mechanical Systems
Smaller component size and cost
Embeddable Microsensors
6
Example early adopter applications CENS Systems
under design/construction

• Biology/Ecosystems
• Microclimate monitoring
• Triggered image capture
• Canopy-net (Wind River Canopy Crane Site)
• Contaminant Transport
• County of Los Angeles Sanitation Districts
  (CLASD) wastewater recycling project, Palmdale,
  CA
• Seismic monitoring
• 50 node ad hoc, wireless, multi-hop seismic
  network
• Structure response in USGS-instrumented Factor
  Building w/ augmented wireless sensors

7
Systems Challenges and Services

• Resource constrained nodes (energy, comm,
  storage, cpu)
• Irregular deployment and environment
• Dynamic network topology
• Hand configuration will fail
• Scale, variability, maintenance

Localization Time Synchronization
Calibration

• Routing and transport in a Tiered architecture
• Channel/connectivity characterization
• Time synchronization and Localization services
• In Network Processing
• Programming model

In Network Processing
Programming Model
Event Detection
8
Tiered Architecture for scalability, longevity

• One size does not fit all.Combine heterogeneous
  devices as in memory hierarchies
• Small battery powered Motes (Mica2 8 bit
  microcontrollers, TOS, 10s of Kbps, 600kbytes
  storage) hosting in situ sensors
• Larger solar powered Microservers (32-bit
  processors, linux OS, 10s of Mbps, 100 Mbytes
  storage)
• Data centric routing/transport at both levels
• Pub/sub bus over 802.11 to Databases,
  visualization, analysis
• Tinydiffusion multihop transport, tasking over
  duty-cycling MAC

9
Network Architecture Can we adapt Internet
protocols and end to end architecture?

• Internet routes data using IP Addresses in
  Packets and Lookup tables in routers
• Humans get data by naming data to a search
  engine
• Many levels of indirection between name and IP
  address
• Works well for the Internet, and for support of
  Person-to-Person communication
• Embedded, energy-constrained (un-tethered,
  small-form-factor), unattended systems cant
  tolerate communication overhead of indirection

10
Data Centric Routing

• Data centric approach has the right scaling
  properties
• name data (not nodes) with externally relevant
  attributes (data type, time, location of node,
  SNR, etc)
• diffuse requests and responses across network
  using application driven routing (e.g., geo
  sensitive)
• support in-network aggregation and processing
• Not just end to end data delivery
• Not just a database query--storage is also
  constrained

Average Packet/Energy Overhead
Communication Radius
Heidemann et.al. SOSP 01, Krishnamachari et
al. 02
11
Diffusion One Phase Pull
Sources
Sink

• Optimized version of general diffusion (Heidemann
  et al.)
• Pulls data out to only one sink at a time (saves
  energy)
• Used in Ecosystem application over Mica 2
  motesTinyDiffusion (Osterweil et al,
  Heidemann-etal04)

Interest
Routed Data
Gradient
12
Voronoi Scoping Restricted Floods from Multiple
Sinks

• Benefits of multiple sinks
• Reduce average path length
• Equalize load over multiple trees
• Tiered architecture, redundancy
• BUT Linear increase in interests flooded!
• Voronoi clusters partition topology, each subset
  contains nodes closest to associated sink.
• Only fwd interests from closest sink
• No overlap between floods
• Motes receive interest from their closest sink
• Scalable both tiers grow, load per mote remains
  constant.

• Live network (emstar/emview)
• 3 sinks, 55 motes
• color-coded clusters

With Henri Dubois-Ferrière, EPFL
13
Multi-hop data extraction characteristics using
Tiny Diffusion

• Collected basic network characteristics to
  verify readiness for sensor deployment (May 03)
• Average system loss rates analyzed over fixed
  intervals and related to nodes of with various
  average, minimum, and maximum hop counts (under
  3 end to end)
• Additional nodes deployed to augment persistent
  ESS topology to study effects such as loss
  experienced by nodes introduced with less ground
  clearance.
• UCB/Intel GDI deployment has good results from
  their fielded borrow monitoring system using same
  Mote platform

14
Characterizing wireless channels

• Great variability over distance (50-80 of
  communication range, vertical lines).
• Reception rate not normally distributed around
  mean and standard deviation.
• Real communication channel is not circular.
• 5 to 30 asymmetric links.
• Not correlated with distance or transmission
  power.
• Primary cause differences in hardware
  calibration (rx sensitivity, energy levels,
  etc.).
• Time variations correlated to mean reception
  rate, not distance from transmitter.

Asymmetry vs. Power
Reception v. Distance
Standard Deviation v. Reception rate
What Robert Poor (Ember) calls The good, the
bad and the ugly
Cerpa, Busek et. al
15
The Ceiling Array A Real Wireless Channel
Motes used to transmit and receive packets -- A
real-world augmentation to a virtual simulation
16
Time Synchronization Service

• Also crucial in many other contexts
• Ranging, tracking, beamforming, security, MAC,
  aggregation etc.
• Global time not always needed
• NTP often not accurate or flexible enough
  diverse requirements!
• New ideas
• Local timescales
• Receiver-receiver sync
• Multihop time translation
• Post-facto sync
• Mote implementation
• 10 ?s single hop
• Error grows slowly over hops
• Elson et al

17
Localization of Sensor Nodes

• Robust ranging
• Wideband acoustics
• Scalable distributed algorithms
• Collaborative multilateration(with beacons)
• Geometry-driven beacon-less
• Fundamental error analysis
• Cramer-Rao bounds for multihop
• Geometry effects
• Angle vs. distance
• Implementation
• MK-II platform with ultrasound ranging
• IPAQs, Mica2s
• Srivastava, Saavedes et al

18
Calibration, or lack thereof
Un-calibrated Sensors

• Storage, forwarding, aggregation, triggering
  useless unless data values calibrated
• Calibration correcting systematic errors
• Sources of error noise, systematic
• Causes manufacturing, environment, age, crud
• Traditional in-factory calibration not sufficient
• must account for coupling of sensors to
  environment
• Nearer term identify faulty sensors and flag
  data, discard for in-network processing
• Significant concern that faulty sensors can wreak
  havoc on in network processing

72º
Factory Calibrated Sensors T0
72º
72º
72º
72º
72º
72º
Factory Calibrated Sensors Later
62º
70º
72º
71º
72º
72º
Dust
Bychkovskiy , Megerian, Potkonjak
19
Information Theoretic FoundationsScalability
for Point Sources in Sensor Networks

• Information theory concerned with fundamental
  limits
• Capacity maximum reliable communication rate,
  versus bandwidth and power
• Rate-Distortion minimum required rate to
  describe source, subject to distortion constraint
  
• Gupta Kumar early result showed wireless
  communication networks do not scale with node
  density per node capacity goes to zero
• However Cooperative rate distortion coding
  results in most communication being local in
  sensor networks-- more nodes do not necessarily
  result in more traffic
• More relays enable increased frequency re-use
  capacity can increase without bound
• Pottie et al

20
Information Theoretic Foundations Scalability
for Distributed Sources

• To estimate parameters of a field (e.g., to get
  isotherm map) information increases until achieve
  desired spatial sampling
• After this extra nodes contribute no additional
  information, but can increase communication
  resource
• Image processing analogy specify pixel size
• Parameters to describe local field can be compact
  compared to raw data, for given level of
  distortion
• Pottie et al

21
Information Theoretic Foundations Practical
Implementation

• Dense network in neighborhood have mix of nodes
  with different ranges, operating in separate
  bands
• Perform local fusion to avoid long-range
  communication of raw data
• Locally route towards the longer range links
  they act as traffic attractors, causing number of
  hops at any given layer to be small, limiting
  delay
• Result is a (largely) standard overlay
  hierarchical network
• Pottie et al

22
In Network ProcessingDistributed
Representation, Storage, Processing

• In network interpretation of spatially
  distributed data
• Statistical or model based filtering
• In network event detection and reporting
• Direct queries towards nodes with relevant data
• Trigger autonomous behavior based on events
• Expensive operations high end sensors or
  sampling
• Robotic sensing, sampling
• Support for Pattern-Triggered Data Collection
• Multi-resolution data storage and retrieval
• Index data for easy temporal and spatial
  searching
• Spatial and temporal pattern matching
• Trigger in terms of global statistics (e.g.,
  distribution)
• Exploit tiered architectures

23
Tiered Data Processing

• Processing uses a two tiered network.
• Task divided into local computation and cluster
  head computation.
• Scope of local computation depends on relative
  cost of local (blue-blue) and cluster-head
  (blue-red) communication
• Example identify regions over which large
  gradient occurring
• Locally, large gradients detected and traversed
  (up to some scope)
• Gradients paths over length threshold identified
  and reported
• Each cluster head combines identification results
  and classifies

T. Schoellhammer, et al
24
Networked Info Mechanical Systems (NIMS)

• NIMS Architecture Robotic, aerial access to full
  3-D environment
• Enable sample acquisition
• Coordinated Mobility
• Enables self-awareness of Sensing Uncertainty
• Sensor Diversity
• Diversity in sensing resources, locations,
  perspectives, topologies
• Enable reconfiguration to reduce uncertainty and
  calibrate
• NIMS Infrastructure
• Enables speed, efficiency
• Low-uncertainty mobility
• Provides resource transport for sustainable
  presence
• (Kaiser, Pottie, Estrin, Srivastava, Sukhatme,
  Villasenor)

25
Actuation NIMS Adaptive Diversity
26
Em Software environment for developing and
deploying wireless sensor networks
Collaborative Sensor Processing Application
Domain Knowledge
3d Multi- Lateration
State Sync
Reusable Software
(Flexible Interconnects not a strict stack)
Topology Discovery
Acoustic Ranging
Neighbor Discovery
Reliable Unicast
Leader Election
Time Sync
Radio
Sensors
Audio
Hardware
27
Em Supports A Slow Descent into Reality

• EmStar allows the same Linux code to be used
• In a pure (low-fidelity) simulation
• Mostly simulated, but using a real wireless
  channel
• In a real testbed, small-scale but
  high-visibility
• Deployed, in-situ, at scale -- but low
  visibility
• Advantage over traditional simulators the
  debugged code itself, not just the high-level
  concepts, flow from simulation into the real
  world
• To maintain high visibility, we trade scale for
  reality

28
Research ChallengeMacroprogramming

• How to specify what, where and when?
• data modality and representation,
  spatial/temporal resolution, frequency, and
  extent
• How to describe desired processing?
• Aggregation, Interpolation, Model parameters
• Triggering across modalities and nodes
• Adaptive sampling
• Primitives
• Annotated topology/resource discovery
• Region identification and characterization
• Intra-region coordination/synch
• System health data, alerts
• Topology, Resources (energy, link, storage)
• Sensor data management (buffering, timing)

Greenstein, Culler, Kohler
29
Lessons

• Channel models
• Simplistic circular channel models can be very
  deceiving so experimentation and emulation are
  critical
• Named data
• Is the right model but its only a small step
  toward the bigger problem of in-network
  processing, macroprogramming
• Duty cycling
• Critical from the outsetand tricky to get
  right--granularity, level (application or
  communication)
• Tiered Architectures
• Optimal distribution of resources (energy,
  storage, comm,cpu) across the distributed system
  is an interesting problem
• NIMS provides an exciting/powerful tier in system
  ecology
• Systems need to be programmable/taskable

30
New Directions
Security
Precision Agriculture
Global seismic Grids/facilities
Tropical biology
Theatre,Film,TV
Coral reef
Macro-Programming
Adaptive Sampling
High Integrity ENS
RFIDs
NIMS
31
Ethical, Legal, Social Implications
Pervasive Computing The ethical, legal, and
policy issues must be addressed during the design
and use stages of these Embedded Network
systemsA more in-depth analysis of public
policy issues is urgently needed that would lead
to appropriate recommendations for solving likely
problems.National Academy of Sciences

• D. Cuff, J. Kang
• Interesting Developments
• RFIDs You might not care about someone
  tracking your razor bladesbut what about your
  tires? (Jay Warrior, Agilent)
• Camera phones
• Fusion of sensor modalities

32
Follow up

• Embedded Everywhere A Research Agenda for
  Networked Systems of Embedded Computers, Computer
  Science and Telecommunications Board, National
  Research Council - Washington, D.C.,
  http//www.cstb.org/
• Conferences ACM Sensys (Nov 03), WSNA, IPSN,
  SNPA (ICC), Mobihoc, Mobicom, Mobisys, Sigcomm,
  Infocom, SOSP, OSDI, ASPLOS, ICASSP,
• CENS website http//cens.ucla.edu (posters from
  recent research review)
• Whose involved
• Active research programs in many CS (networking,
  databases, systems, theory, languages) and EE
  (low power, signal processing, comm, information
  theory) departments
• Industrial research activities at Intel, PARC,
  Sun, HP, Agilent, Motorola
• Startup activity at Crossbow, Sensicast, Dust
  Inc, Ember,
• Related Funding Programs
• DARPA SenseIT, NEST
• NSF ITR, Sensors and sensor networks