Growth Codes: Maximizing Sensor Network Data Persistence

1
Growth CodesMaximizing Sensor Network Data
Persistence

• Abhinav Kamra, Vishal Misra, Dan Rubenstein
• Department of Computer Science, Columbia
  University

Jon Feldman Google Labs
2
Outline

• Problem Description
• Solution Approach Growth Codes
• Experiments and Simulations
• Conclusions and Ongoing work

3
Background A generic sensor network
Sensor Nodes
Data follows multi-hop path to sink(s)
Sink(s)
x1
x9
Sensed Data
x10
x2
A single node failure can break the data flow
x12
x11
x3
x8
x5
x6
Generic Aim Collect data from all nodes at
sink(s)
x13
x7
x4
4
Specific Context for our problem

• Sensor Networks in a Disaster setting
• E.g., Monitoring earthquakes, fires, floods
• Problems in this setting
• Congestion near sink(s)
• All nodes simultaneously forward data
• Overwhelm sink(s) capacity
• Network Collapsing nodes failing rapidly
• Pre-computed routes may fail
• Data from failed nodes can be lost

5
Challenges

• Networking Challenges
• Disaster scenarios feedback often infeasible
• Frequent disruptions to routing tree if setup
• Difficult to predict node failures sink
  locations unknown, surviving routes unknown
• Difficult to synchronize clocks amongst nodes
• Coding Challenges
• Data source distributed (among all sensor nodes)
• Prior approaches (Turbo codes, LDPC codes) aim at
  fast complete recovery
• Sensor nodes have very limited memory, CPU,
  bandwidth

6
Objectives

• Fraction of data that eventually reaches the
  sink(s)

Preserve data from failed sensor nodes

6 of 10 symbols reach sink. Persistence 60
Deliver data to sink(s) as fast as possible

Maximize Data Persistence
7
Limitations of Previous Work

• Channel Coding based
• (e.g. Turbo Codes Anderson-ISIT94, LT Codes
  Luby02)
• Aim for complete recovery in minimum time
• Difficult to implement with distributed sources
• Routing-based
• (e.g. Directed Diffusion Govindan00, Cougar
  Yao-SIGMOD02)
• Conjecture Is fragile (disrupted easily) in
  disaster scenarios

8
Our Approach

• Two main ideas
• Randomized routing and replication
• Avoid actively maintaining routes
• Replicate data to increase data survival
• Distributed channel codes (Growth Codes)
• Expedite data delivery survivability

First (to our knowledge) distributed channel codes
9
Outline

• Problem Description
• Our Solution Growth Codes
• Experiments and Simulations
• Conclusions and Ongoing work

10
Network Assumptions
4
3
2
5
S
1
6
7
S

• N node sensor network
• Limited storage each node stores C data units
• Large storage at sink(s)
• All sensed data assumed independent (no source
  coding)

11
High Level View of the Protocol
4
1
2
3
Nodes send data at random times (Current
implementation exponentially distributed timers)
12
High Level View of the Protocol (2)
1
2
?
0
Degree 2 codeword
Even if node 3 fails Node 3s data survives
K1
Sender picks a random symbol
XORs it with its own symbol
K2
K3
After time K1, nodes start sending degree 2
codewords
13
High Level View of the Protocol (3)

• After time K1, nodes start sending degree 2
  codewords
• After time K2, nodes start sending degree 3
  codewords
• 
  .
• 
  .
• 
  .
• After time Ki, nodes start sending degree i1
  codewords

Times Ki can be out of sync at different nodes
No need to tightly synchronize clocks
14
The Intuition behind Growth Codes
Codewords
When very few symbols decoded
Easy to decode low degree codewords
Set of symbols decoded at Sink
time line
15
The Intuition behind Growth Codes(2)
Codewords
When significant number of symbols decoded
Low degree codewords often redundant
Set of symbols decoded at Sink
Higher degree codewords more likely to be useful
16
Outline

• Problem Description
• Growth Codes
• Experiments and Simulations
• Conclusions and Ongoing work

17
Simulations/ExperimentsCompare data persistence
of various approaches

• Simulations
• Centralized Setting compare GC with other
  channel coding schemes
• (e.g. Soliton, Robust Soliton) LT Codes
  Luby02
• Distributed Simulation to assess the performance
  gain of coding vs no coding
• Experiment on motes
• compare time for complete recovery of GC vs
  routing
• resilience to node failures

18
Comparison with various coding schemes(N 1500)

• Centralized Simulation
• (to compare with other channel coding schemes)
• Single source, single sink
• Source generates codewords according to coding
  scheme (GC, Soliton, R-Soliton)
• Zero failure rate

• No coding is fast in beginning
• Slowdown is explained via Coupon Collectors
  problem
• Soliton/ R-Soliton slow in the beginning (use lot
  more higher degree codewords)
• Growth Codes tries to decode the maximum at any
  time

1
19
Growth Codes vs No Coding(Varying N)

• Distributed Simulation
• (to assess the performance gain of coding)
• N sources, single sink
• Random graph topology (avg degree 10)
• Sink receives 1 codeword per time unit

• Complete recovery takes
• O(N logN) time without coding (Coupon Collectors
  effect)
• Linear time with Growth Codes
• Soliton/R-Soliton have no distributed
  implementation. How to compare?

20
Experiments with (micaz) motes

• (to measure data persistence with time)
• GC vs TinyOSs MultiHop routing protocol
• No routing setup at beginning (scenarios where
  sensor nodes are deployed rapidly)

• MultiHop for persistence takes long time to
  complete route setup
• Comparison with GC simulator validates simulator
  performance

21
Motes experimentsResilience to node failures

• Nodes generate data every 300 seconds
• 3 nodes fail just after 3rd data generation

22
Motes experimentsResilience to node failures

• 1st generation, GC faster, MH takes time to setup
  routes
• 2nd generation, routing already setup, MH very
  fast
• 3rd generation, MH needs to repair routes

23
Conclusions

• Data persistence in sensor networks
• First distributed channel codes (GC)
• Protocol requires minimal configuration
• Is robust to node failures
• Simulations and experiments on micaz motes show
• GC achieves complete recovery faster
• GC recovers more partial data at any time

24
Ongoing Work

• Adapt Growth Codes to scenarios where sensor data
  is correlated
• Take advantage of any available routing
  information (e.g. before a disaster)
• Estimate network size on the fly to use in Growth
  Codes

25
Thanks for your patience !

• For more information
• DNA Research Lab, Columbia University
• http//dna-wsl.cs.columbia.edu/