Data Dissemination with Geometric Structure

1
Data Dissemination with Geometric Structure

• Gilman Tolle
• UC Berkeley

2
Overview

• Deluge A bulk data dissemination system
  supporting multihop in-network reprogramming
• Simulation of Deluge in large networks shows that
  increasing radio neighborhood size results in
• excessive hidden-terminal collisions
• under-suppression and extra contention
• much slower data propagation
• Deluge protocol is geometry-independent,
  selecting senders based on reception of
  advertisements
• Research question Can modifying Deluge to use
  geometric information lessen the negative effects
  of density?

3
Geometric Background

• The tested optimizations assume that each node
  has perfect knowledge of the location of each
  other node, and of the link qualities between
  them
• Too idealistic? May help to establish a lower
  bound on dissemination performance
• TOSSIM simulation of 400-node grid
• Grid chosen to support creation of geometric
  patterns
• Geometric location and radio connectivity are
  well-correlated in TOSSIM
• Basic metric propagation time

4
Geometric Propagation Patterns

• Line
• Create artificial edges through center
• Disseminate to contended areas first
• Tree
• Fixed parent chosen from optimal MET tree
• Intended for small data, good for large data?
• Fractal
• Recursively subdivide network into edges
• These approaches will lessen parallelism, but may
  improve contention

5
Neighborhood Knowledge

• Density Awareness
• Each node is aware of the size of its
  neighborhood
• Proactively increase backoff times
• Tried both linear and exponential functions
• Neighborhood Reduction
• Restrict possible senders to immediate neighbors
• Encourages stronger links

6
Line Results

• Balance time spent in line with open flooding
• Fixed parent requirement is very sensitive to
  poor link qualities
• Line to center was faster (10)
• Routing data to center is slightly beneficial

7
Tree Results

• Maintains speed through center region
• 10 speedup
• Also sensitive to links, but
• Chosen sender is well-connected
• Does not consider sibling sender spacing

8
Density Awareness Results

• Explicit backoff does prevent tendency to
  under-suppress
• Lessens contention between neighboring requests
  and improves request reception rate
• 10 speedup, but very sensitive to constants
  used in function
• Additional benefit could be used in conjunction
  with other techniques

9
Neighborhood Reduction Results

• Possible senders are restricted to 4 Manhattan
  neighbors, then 8 neighbors including diagonals
• Overall best speedup (25)
• Prevents selection of a poorly-connected sender
  based on one long-link advertisement
• Allows flexibility in determining sender spacing
  that single-parent patterns cannot provide
• Suggests that history-free sticky protocols are
  vulnerable to inadvertent poor sender selection

10
Contention Metrics
Condition Loss-Independent Reception Rate (Mean) Request to Data Ratio (Mean)
Baseline 0.6 1.17
Line-Center 0.51 1.61
Tree 0.72 0.63
8-Nbrs 0.74 0.4

• Tree and Neighborhood Reduction increase
  reception rate and decrease number of required
  requests
• Line actually shows worse contention
• Multiple senders, or starting in dense region

11
Flexibility Tradeoffs

• All patterns depend on foreknowledge, but to
  varying degrees
• Line and Fractal
• global coordinates for nodes
• Density Awareness
• needs actual neighborhood size
• Tree and Neighborhood Reduction
• known link qualities

12
Geometry and Estimation

• Most techniques can also be based on estimation
  done in advance
• Line and Tree
• Build a routing tree
• Select destination node
• Route along the path
• Density Awareness
• Count neighbors over time
• Neighborhood Reduction
• Estimate neighborhood link qualities
• Estimating in advance assumes that geometry of
  network is constant during dissemination...valid?

13
Simulation Sensitivity

• Interference model in TOSSIM is overly
  pessimistic
• Packet success rate does fall off with distance,
  but...
• Every node within 50 feet has an equal chance of
  causing a collision
• Testing with 25-foot interference range results
  in 3x speedup using basic Deluge protocol
• Optimizations help less
• Neighborhood reduction is only technique that
  improves performance, and only by 20

14
Contention Metrics (25-foot)
Condition Loss-Independent Reception Rate (Mean) Request to Data Ratio (Mean)
Baseline 0.85 (vs. 0.6) 0.45 (vs 1.17)
Line-Center 0.82 (vs. 0.51) 0.44 (vs. 1.61)
Tree 0.98 (vs. 0.72) 0.13 (vs. 0.63)
8-Nbrs 0.90 (vs. 0.74) 0.15 (vs. 0.4)

• All contention metrics improve over 50-foot
  interference
• Tree and Neighborhood Reduction still improve
  over baseline

15
Overall Performance

• Could obtain no more than a 25 speedup from any
  particular technique
• Geometry-independent Deluge may be close to
  optimal
• Real-world results likely to be even less
  dramatic
• Estimated geographic information
• Weaker interference as compared to simulator

All, w/25 ft Interference
Line Patterns
Tree, Density, Nbrs
16
Conclusions

• Even perfect geographic information only improves
  speed by up to 25, in simulation
• Comes at the expense of assuming a static network
  organization and resulting loss of flexibility
• Overly pessimistic interference model in TOSSIM
  suggests that previously observed contention
  effects may be exaggerated
• Future work establish tight lower bound,
  simulator revalidation