Geographic Routing Made Practical

1
Geographic Routing Made Practical

• Young-Jin Kim, Ramesh Govindan, Brad Karp, Scott
  Shenker
• Presented by Jorge Ortiz

2
Scalability in Routing
Internet Routing (Distance-Vector,
Link-State) Routing table size O(N) per node For
correctness, routing table must be completely
up-to-date Geographic Routing A scalable class
of routing algorithms for wireless
networks Routing table size O(d) per node,
number of single-hop neighbors but exclusively
evaluated in simulation

• ? Does geographic routing work correctly in real
  world?
• ? irregular radio range
• ? radio-opaque obstacles
• ? imperfect localization

3
Introduction to geographic routing
1. Greedy traversal G.G. Finn 87

• - Nodes learn 1-hop neighbors positions from
  beaconing

• A node forwards packets to its neighbor closest
  to D

4
Introduction to geographic routing
Greedy traversal not always possible!
x is a local minimum to D w and y are far from D
5
Introduction to geographic routing
2. Face (Perimeter) traversal The Right hand
rule

• Well-known graph traversal right-hand rule
• Requires only neighbors positions

z
y
x
Fails when there are cross links in the graph!
6
Introduction to geographic routing
2. Face (Perimeter) traversal on a planar graph
Two primitives (1) the right-hand rule
(2) face-changes
D
F4
F3
F2
Walking sequence F1 -gt F2 -gt F3 -gt F4
F1
X
? Many existing algorithms like GFG, GPSR,
GOAFR, and etc. combine greedy traversal with
face traversal.
7
Introduction to geographic routing
3. Planarization techniques GG or RNG

• Given a radio graph, make a planar sub-graph in
  which every cross-edge is eliminated.

• ? Important assumptions
• - Unit-disk graph Accurate localization
• ? How well do planarization techniques work
  in real-world?

w
u
v
?
Gabriel Graph
GG (Gabriel Graph)
GG Sub-graph
Full Radio Graph
8
GPSR in network test-beds
? A test-bed deployed in UC Berkeley Soda Hall
50 MICA2dot, 433MHz radio, 5.2 average node
density
Cross-link

• 68.2 routing success among node pairs
• Whats happening?

Unidirectional
Disconnected
Wireless Network Graph
GG sub-graph
9
Mutual Witness

• ? Key idea
• - remove a link only if both ends of the
  link see a mutual witness
• - can eliminate unidirectional links,
  disconnections

• Raises success rate to 87
• But, mutual witness introduces other failure
  modes
• - converts unidirectional/disconnected links into
  cross links
• leaves cross links in a sub-graph
• generates collinear links (a degenerate case)

10
CLDP (Cross-Link Detection Protocol)

• Basic Idea
• - Each node probes its links to determine
  crossed links.

• ? Key features
• 1. Completely distributed protocol
• - Each node executes this protocol on all
  links
• 2. Can prove that, when CLDP executes on any
  arbitrary graph, face traversal never fails on
  the resulting sub-graph

pcrossings of (S, A)?
B
A
p(B, C) crosses (S, A)!
pcrossings of (S, A)?
S
C
p(B, C) crosses (S, A)!
11
Avoiding network partition
Network partition!
Cross links!
B
A
A
B
C
C
S
S
case A
case B
Solution ? keep a link if the probe returns on
the link it was sent out on ? can leave
cross-links in the sub-graph, but our proof shows
that face traversal cannot fail on the
sub-graph
12
Avoiding race conditions

• Problem
• Concurrent CLDP probing causes race
  conditions

Network partition by race condition!
Solution ? lazy locking mechanism ?
described in the paper
13

• Experimental Results Simulation
• Radio graphs with obstacles
• Random graphs
• Metric
• Success rate
• Stretch
• Overhead
• Convergence

14
Success Rate

• TOSSIM
• 200 nodes

Radio graphs with 200 obstacles

• Observations
• CLDP is perfect on radio graphs with obstacles.
• GPSR performs poorly due to partitions and
  unidirectional links

15
Success Rate
Random graphs

• Observations
• Even on random graphs, CLDP is perfect.

16
Stretch

• measured path length / shortest path length

Radio graph
CLDP subgraph
GG subgraph

• Observations
• CLDP (24) outperforms GPSRGG/MW

- GG planarization removes more links than CLDP
17
Stretch distribution
Radio graphs with 200 obstacles

• Observations
• Worst-case stretch - 102

Geographic routing uses only local information
to determine paths!
18
Worst case stretch
? Worst case stretch O(n), n number of
nodes ? Worst case path length O(nl), l
optimal path length ? Recently, we have
discovered techniques that can improve
worst-case behavior
S
D
Right-hand rule
Example from test-bed experiment
19
Experimental Results Implementation

• nesC Code (about 4500 lines)
• Run on MICA2 (CC1000 Radio)

20
Experiments on test-beds

• Two sensor network test-beds (Berkeley-433MHz
  and Intel-916MHz)

1. Rm 50 nodes on Berkeley Soda Mica2dot
test-bed 2. Rs 25 nodes on Berkeley Soda
Mica2dot test-bed 3. C 36 nodes on Intel
Berkeley Mica2dot test-bed

• Success rate of all source-destination pairs.

• CLDP is immune to all pathologies

21
Stretch

• measured path length / shortest path length

• Mostly (8197) below 3
• But, worst case 20

22
Overhead

• the number of CLDP probes observed on each link
• during a probing interval (15 seconds in our
  implementation)

• low overhead below 7 for 90 of links
• depends on the size of faces and the no. of
  cross links
• affected by the deployment node density

23
Steady-state Convergence
v Over 80 of links converge within 3 probing
intervals v Almost all links converge within 6
probing intervals.
24
Conclusion

• Geographic routings GG planarization suffers
  from 3 pathologies.
• asymmetric links, network partition, and
  cross links
• Augmenting GG planarization with MW is not
  enough.
• other illness such as cross-links and
  collinear links
• CLDP
• ? the first distributed planarization
  protocol
• - immune to pathologies that cause persistent
  routing failures on static networks.
• - thus, guarantees routing success for all
  node pairs.
• ? shows good average stretch, reasonable
  overhead and
• convergence time.