Partition Detection

1
Partition Detection

• EDIFY Overview
• Brian D. Davison

2
Partition Detection Papers

• Localized algorithms for detection of critical
  nodes and links for connectivity in ad hoc
  networks
• Jorgic, Stojmenovic, Hauspie, and Simplot-Ryl,
  3rd IFIP MED-HOC-NET Workshop, 2004
• Partition Detection in Mobile Ad-Hoc Networks
  Using Multiple Disjoint Paths Set
• Hauspie, Carle, and Simplot, Objects, Models
  Multimedia Tech. Workshop, 2003
• A Partition Detection System for Mobile Ad-Hoc
  Networks
• Ritter, Winder, and Schiller, IEEE SECON 2004

3
Localized algorithms for detection of critical
nodes and links

• Jorgic, Stojmenovic, Hauspie, and Simplot-Ryl,
  IEEE MASS, 2004
• Recognizes that before partitioning, there are
  critical nodes and links
• A node or link is critical if the subgraph of
  k-hop neighbors of the node is disconnected
  without the node
• Once critical nodes/edges are known
• services or data may be replicated before a
  partition occurs
• Node trajectories could be changed to delay or
  prevent partitioning (perhaps by re-inforcing the
  critical link) which will directly increase
  delivery rates
• DFS was used to detect critical links
• Global algorithm
• Can be centralized or distributed
• Communication costs are expensive
• Possible partition can be detected with high
  probability with localized algorithms
• Reduced communication overhead
• Increased detection speed

4
Literature

• Shah, Chen, and Nahrstedt (SCI 2001)
• Used GPS to monitor position, computes velocity,
  can predict when partitioning will occur
• Wang and Li (INFOCOM, ICC 2002)
• Detected future partitions, replicated services
• Strong centralized approach
• Hauspie, Simplot and Carle (Med-Hoc 2003
  Workshop)
• Evaluated stability of path from source to
  destination with function of disjoint path
  between them and hop distances
• Significant communication overhead to evaluate
  function

5
Localized partition detection

• Localized algorithms
• More scalable, robust, and energy efficient
• May (falsely) detect some nodes as critical
• May still be better to replicate data/services
  when c/s is far apart anyway
• Definitions
• Local (k-hop) knowledge
• k-hop neighbors shortest route is k or less hops
• Collected by sending hello messages to neighbors
  containing graph of their k-1 hop neighbors
• Positional vs. topological information
• (positional tells about connections between nodes
  2 hops away, while topological does not)
• Topological is never better than Positional

6
Localized algorithms

• Critical node detection
• A node is k-critical if the subgraph of k-hop
  neighbors is disconnected
• If globally critical, it will be detected
• Critical link detection
• Link AB is critical if the sets of k-hop
  neighbors of A and B are disjoint
• Detection of critical links based on loop length
• Link UV is k-loop-critical if hop distance
  between U and V is gt k (after UV is removed)
• Detection of critical links based on critical
  nodes
• Link AB is k-critical if both A and B are
  declared as k-critical nodes
• Much less communication required (just k-critical
  status)

7
Performance evaluation

• Test algorithms on connected random graphs with n
  nodes, average degree d
• N100, 500
• Densities of 3-15
• Measure is detection ratio
• Probability that a node/link is declared as
  critical by local algorithm is indeed globally
  critical
• Also measured average number of critical nodes
  and links detected

8
Detection ratios
9
Ave number of critical nodes
10
Conclusions

• Localized algorithms give excellent results
• Detection of critical nodes is about as accurate
  as detection of critical links
• Difficulty is in detection of ring structures
  longer than k
• Little relation to density

11
A Partition Detection System for Mobile Ad-Hoc
Networks

• Ritter, Winder, and Schiller, IEEE SECON 2004
• Proposes two partition detection schemes
• Centralized
• Distributed
• Motivating example mobile multiplayer games
• Group moves and partitions occur

12
Related work

• Babaoglu et al., Bacon et al. partition-aware
  apps
• can reconfigure themselves after a merge
• Uses simple ping/ack mechanism with timeout
• Cannot distinguish between node failure and
  partitioning
• Killijian et al.
• Failure anticipation, movement planner, etc.
• Expensive

13
Partition detection

• Two sets of nodes either active participants in
  monitoring system, or not
• Active (probing) nodes
• must be chosen carefully to ensure most of the
  topology is monitored
• Have relatively degree (closer to the border of
  the network)
• Periodic keep-alive messages between active nodes
  (far apart)
• When not heard after some time period, partition
  suspected
• Local validation also use one-hop neighbor to
  watch buddy, and notify other monitoring node if
  buddy fails

14
Active (border) node identification

• Nodes with small enough degree
• Static approach uses fixed threshold (degree lt
  threshold)
• Dynamic approach sets threshold as last set of
  neighbor counts received from neighbor
• Because of fluctuations in mobile networks, use a
  hysteresis with different thresholds to change
  state

15
Centralized Partition Detection

• One node sends beacon messages
• First active elects itself, broadcast to rest
• Other nodes are recipients
• Disadvantages
• partition containing server does not detect the
  partition
• Partitionings containing no active nodes will not
  be detected

16
Distributed Partition Detection

• Every active node sends beacons
• Broadcast on becoming active
• Every node caches a set of recent (and/or far
  away) node announcements, uses them as partner
  nodes
• All active nodes need a buddy node
• Buddy is told of partner node list changes

17
Experimental Results

• Implemented both in ns2
• 50 nodes, simulated 50 times for each parameter
  combination
• Centralized approach
• Low message overhead
• Network unmonitored after server failure but
  before active nodes register at new server
• Distributed approach
• required 7x the number of messages messages
  spread across many nodes
• much more resilient to node failure