1. Self-Organizing Hierarchical Routing for Scalable Ad Hoc Networking
At Ad/Hoc Networks, Elsevier, 2007
Or Computer Science Technical Report
Rice University, 2004
Slides by Ece Gelal and M. Faloutsos

2. Problem Use of wireless devices is rapidly increasing
Decentralized, self organizing systems
Current ad hoc routing protocols are limited in scalability
Mobility increases routing overhead
Success of the protocols is challenged in large networks
Need for routing protocols that scale

3. Contribution Self-organizing hierarchical routing
Proximity-based multi-level hierarchy of cells
Address = concatenation of cluster ids
Hybrid Routing using this hierarchy
Proactive hierarchical, among cells
On-demand, within cell
Mapping of unique IDs to hierarchical addresses using a distributed hash table
Nodes make individual decisions to provide and preserve the hierarchy

4. The Main Idea
Level
Each node belongs to a cluster of level L, or is clusterhead (Drum) for that level
Drums broadcast their existence with beacons
Nodes hear beacons and
Join cluster
Take over drum role 3 2 1

5. DART versus SAFARI What is common? Difference from DART
Hierarchical addressing
Map IDs to addresses via a hash table
Difference from DART
Explicit group names (instead of prefixes)
Reactive Routing within same level-1 cell

6. Safari Overview: Hierarchy
Nodes organized in multi-level cells
Level k cells grouped into Level k+1 cells
Cells are identified by drums (= head)
Automatic self-selected
Drums are the backbone of the hierarchy
Nodes choose a 1-level higher drum
The one closest in hops

7. Safari Overview: Routing
Self-organize into clusters:
Drums transmit periodic beacons
Nodes keep DHTs (distr. hash tables)
Look up ID of node to find its address
Route towards the drum of the cell of dest.
Up: Follow the reverse path of most recent beacon
Down: Seek a lower-level cell s.t. dest. is its member
Within: Any node in cell of dest. initiates on-demand route discovery

8. Beaconing Protocol Beacons help establish the drums
Drums broadcast beacons periodically:
Seq #, coordinate, beacon level, hop count
Forwarded by all nodes within Dn hops
D_n increases with n: higher level, larger cell
Higher level beacons w/ lower freq.
Assumes:Topology changes less frequent at higher level
A beacon of level n transmits beacon of level n in every tn seconds

9. Clusterhead and hierarchy
k-level clusterheads “rule” a D_k radius area
At k+1-level we have fewer clusterheads
D_k
D_k
D_k+1

10. Beacons: Rules of Hierarchy
DART = Dynamic Address Routing Table
A node at level n, must have a leader at level n+1
There is no other n-level node near by
There is a reason why you are at level n
At least two clusters near you at level n-1

11. Drum Level Selection
Algorithm runs at each node, after each change in DART
New entry or timer expiry
Assure DART satisfies certain conditions
If violated, change the node level appropriately
Aids efficiency under dynamic changes
Algorithm runs after each change in DART

12. Membership Algorithm Invoked after drum level selection
Nodes choose a unique drum in 1-higher level (closest drum)
Whose DART entry has not expired
Then each node is assigned a coordinate, based on its drum
Consider Di is a drum, L is a leaf node
COORD(Di)=COORD(PARENT(Di))RAND(b)
COORD(L)= COORD(PARENT(L))
Coordinate is vital in routing

13. The Routing Table: DART
Drum Ad Hoc Routing Table (DART)
Nodes store beacons they forwarded
Store: beacon contents, time_rcvd, Drum_ID
A timer starts with each new entry
When expired,I rethink my membership
drum level selection algorithm and membership algorithm invoked

14. Addressing -- Implications
COORD(Di)=COORD(PARENT(Di))Rand(b)
if b  then Pr{2 nodes get same coordinate} 
COORD(L)= COORD(PARENT(L))
All leafs in a cell have the same coordinate
Drum IDs at each level create a hierarchical address (coordinate) for each node that is a member.
Each node associates itself with the coordinate of its drum.

15. Proposed Routing Protocol
1. Proactive Inter-Cell Routing
Deliver pkts from src to level-1 cell of dest
Drums broadcast info to
All its nodes
All nodes of its sibling cells (same parent at n+1)
By following the reverse path of beacons
2. On-demand (Reactive) Intra-Cell Routing
Deliver pkts to final dest. within level-1 cell
Via any on-demand routing protocol

16. Proactive Inter-Cell Routing
Unlike conventional clustering
Clusterheads suffer bottleneck
Drums need not be along the path at all
Extract coordinate from pkt, Lookup DART
Longest prefix matching between two coordinates
Get the sequence # of the longest prefix match
Compare with sequence # in pkt header
If not worse, replace entry in pkt with DART info

17. Reactive Intra-Cell Routing
Each node, upon receiving a pkt, checks if the dest is in same level-1 cell as itself
Coordinates are the same
If coordinates match, use DSR within cell
If prefix matching = n, then use DSR, intra-cell routing. COORDINATE S ARE THE SAME!

18. Example
Safari Routing overview

19. Address Resolution Senders wish to send pkts to an ID
But protocol forwards to a coordinate
Map nodeID to its current coordinate
Implemented like DHTs
Each node inserts (ID,coordinate) pair at a set of k nodes
Node reinserts whenever its coordinate changes
Any node can lookup the coordinate by hashing the ID
No additional maintenance overhead

20. Address Resolution - Locality?
We don’t want to look up a coordinate in LA from a node in Japan!
Store hash not only in random k nodes; but also in closest k nodes
Insert, with i least significant coordinate components replaced by random hash value
When coordinate changes in i least significant coordinates, reinsert at a subset of nodes only
When A looks up B’s coordinates
Iteratively query all nodes at greater distances
Lookup of a nearby node does not necessitate communication with a distant node

21. Performance Evaluation
Dn= 3 hops
Transmission range = 250 m.
Topology:
fixed density: 50 nodes in 1000m x1000m area
Mobility: random waypoint

22. Scalability PDR versus network size
PDR: fraction of application layer packets successfully rcvd
PDR versus network size

23. Overhead Overhead asymptotically approaches a constant value.
Overhead due to drum floods
Overhead asymptotically approaches a constant value.

24. Path Lengths
Avg # of hops btwn a random src and a random dest.

25. Mobility 1000 nodes. 95% PDR with 100% mobile nodes.
Total overhead almost same as in the static case.

26. DHT Scalability Avg. # of hops to LOOK UP a coordinate in DHT.
Avg. # of hops for all pkts to
STORE a coordinate in DHT.
Avg. # of hops to
LOOK UP a coordinate in DHT.

27. Conclusions Proposed routing algorithm scales well.
Periodic beacons are an overhead.
A lot of patch-like details create complication
Beacon period
Beacon forwarding
Drum level selection: complicated
Coordinate caching in DHTs
Choose a random number, Listen for collision, if there is collision choose another random number