Transcript estrin_scadds-pimtg

Networking Research Review
SENSIT PI Meeting
October 7-8, 1999
Marina Del Rey
SCADDS (ISI/W) -- Estrin
GRASP (UCLA/CS) -- Zhang
DDNC (MIT-LL) -- Van Hook
DSN (UCLA/EE-ISI/E) -- Srivastava
WINS (Sensorweb) -- Kaiser
General Organization
(15 minutes per project)
Brief overview
Detailed progress since last meeting
Short term issues encountered (if any)
New directions, emphases
SCADDS Recent Progress
(PI’s: Deborah Estrin, Ramesh Govindan,
John Heidemann)
Directed diffusion v0 (Intanago)


Initial simulation results
Initial prototype implementation
Experimental platform (Elson, Girod, Kumar,
Raghunath, Zhao)



Linux and short-range radios
Simple hardware assembled to support protocol
experiments--rf sensors, tags (in progress), using cots
radios
Scaffolding for diffusion and application
SCADDS: Ongoing activities
Preparation for use of WINS ng nodes



Detailed discussions of comm API
Investigation of current and planned assembly
mechanisms(FH and TDMA)
Plan to interface ucLinux nodes directly to Sensorweb
hardware--run diffusion algorithms on SENSIT testbed
Algorithm development and evaluation




Directed diffusion design and evaluation (Chalermak
Intanagonnowat)
Adaptive clustering (Satish Kumar)
Timing/Synchronization (Jeremy Elson, Lewis Girod)
Adaptive fidelity (Amit Kumar, Ya Xu)
Directed Diffusion
Version 0.0 of directed diffusion





Multi-path delivery
Distinct information dissemination
Probabilistic forwarding
Normalized gradients
Initial experiments with one source and one
sink per data type
Many other “flavors” of diffusion worth
exploring
Directed Diffusion
Preliminary “Indications”
Overhead

Early indications that average network
overhead (data, power, state) grows linearly
with network size
Overhead per node is constant

Traffic dependent
Energy Dissipation

Low variance of remaining energy across nodes
Indicator of effective load balancing and long
network lifetime
Supplementary: Directed Diffusion
Future work
Study parameter tuning of the model
Cleaner model : Generalization of
reinforcement and interest
Explore additional flavors of diffusion



Redundant information dissemination
Absolute gradients
Multiple sources and multiple sinks per data
type
Port to WINS ng nodes, or interface our
sensor-controller platform to theirs
Adaptive Clustering
Original hypothesis: Adaptive clustering allows
efficient coordination of local interactions
However cluster creation and maintenance can
consume significant energy that has to be
amortized over gains in application function



Soft-state techniques may consume too much energy
at low query rates
Hard-state techniques perform better but adaptation
may be more difficult (work in progress)
Adaptation is too energy inefficient if frequency of
adaptation not properly controlled
Supplementary: Adaptive Clustering:
TDMA Master Election
Master node assigns TDMA slots to slave nodes

Communication between sensors through master to
conserve energy
Master’s radio powered on all the time and
hence consumes more energy than slaves
Adapt master selection based on energy to
improve network lifetime
Supplementary: Adaptive Clustering:
TDMA Master Election
Adaptation also has a cost:



Energy cost of the re-election process
Potential data loss during adaptation
Potential re-organization of neighbor clusters
Change in cluster membership
Re-assignment of TDMA slots
Some Project Issues
Evaluation Platforms


uclinux hardware? which radio?
Better indoor propagation and power models
for use in non-experimental evaluations
Interfaces and APIs


Interface to WINS ng nodes (i.e., real sensor
data and real low-power radio)
Interface to applications
Interaction of diffusion and radio/mac level
behaviors
Supplementary: Development Platform
ucLinux and ucSimm

The Linux
Microcontroller
project uclinux is a
port of the Linux 2.0 to
" Features
systems without a
"
3.5 in x 1 in x 0.25 in, 30pim SIMM
Memory Management
Unit.
"
16Mhz MC68EZ328 DragonBall

Target Systems:
3Com Palm III+TRG
memory board
"
"
8Mb RAM, 4Mb FLASHROM
I/O Interfaces
Other micro-controller
such as MC68K series
" 18 General Purpose I/O pins
ucSimm: specially
designed simm module
" 10Base-T Ethernet (CS8900A)
" Will directly drive a LCD panel 320x240
" RS-232 Serial
" Approx $150 per node
Supplementary: Pros and Cons
 Open Source: GNU
Public Liciense

Good Portability

Potential Applications
Available worldwide
 Simple but Flexible I/O

Radiometrix Transceiver

A/D, D/A converter

Standard serial or 10Based
wired connection
 Low Power Consumption

3.3v low voltage, 63mA 108mA
 Low Price
" Still in pre-mature stage
"
"
"
Limited Extensibility
limited # of I/O pins
No Standard AddrBus or
DataBus
Summary
 Diffusion experiments underway on prototype testbed

Sensors are Librettos or ucSimm running linux with
Radiometrics radio as rf-sensor

Tags provide data (using small form-factor, semiprogrammable radio beacons)

Can be ported or interfaced to WINS ng nodes for SENSIT
demo in 2000
 Other algorithmic work in design and
modeling/simulation phase

Diffusion--characterization, comparisons

Neighbor identification/coordination/synchronization

Clustering

Adaptive fidelity