1. Routing in Intermittently Connected Wireless Networks
Routing in Intermittently Connected Wireless Networks a brief survey on recent works
Joy Ghosh
LANDER

2. Overview A brief introduction to DTN, ICMAN, etc.
Pocket Switched Networks
DTN Routing in Mobility Pattern Space
Practical routing in DTN
Spray and Wait: Routing for ICMAN
Inter Planetary Networks

3. What are DTN, ICMAN ? Delay Tolerant Networks (DTN)
No contemporaneous end-to-end paths
Store-n-forward routing methodology
Links/contacts are often subject to long delays
Packet losses due to buffer overflows
E.g., busses in fixed routes, satellites in orbits
Delay tolerant networking research group (
Intermittently Connected Mobile Ad hoc Networks (ICMAN)
In the realm of DTN and MANET
Focuses more on partially deterministic mobility

4. Pocket Switched Networks and Human Mobility in Conference Environments [1]
Research Issues
Internet connectivity islands – realms of DTN
Dependence on managed infrastructure (DNS, DHCP, centralized servers, etc.)
Effective use of human mobility as well as local/global connectivity
Experimentation
54 participants of IEEE Infocom 2005 carried bluetooth devices to record connectivity statistics
Implications on PSN forwarding algorithms

5. Pocket Switched Networks
Human mobility – double edged sword
(+) increases bandwidth as users store/carry data
(-) unstable forwarding paths, varying reach-ability
Opportunistic networking
Contemporaneous path is restrictive
Per-hop routing
Make use of local connectivity and node mobility
E.g., Data muling, Store-and-haul forwarding
Locally forwarding to nodes with global connectivity
Return path is tricky – Open Problem
Use of both local and global connections makes it robust

6. Personal Devices Always-On, Always Carried
NOTE: is Always-ON a practical assumption?
Support owner’s task before others
If resources can be spared
Security and Privacy
Uncontrolled and potentially malicious neighbors

7. Human mobility measurements
Setup with 54 attendees of Infocom 2005
Intel iMotes with ARM processor, Bluetooth radio, Flash RAM, CR2 battery, packed in a floss box
Bluetooth base band layer “inquiry” mode for 5 secs
Sleep for ~120 seconds, responding to only “inquiry”
Maintain “in-contact” list; when nodes don’t respond, write a tuple {MAC, start time, end time} to flash RAM (64K for data)

8. Contact Visibility
Influence of time of day

9. Contact Duration - I
Contacts between specific pairs of nodes

10. Contact Duration - II
Contacts with any node in a group

11. Inferences
Combination of local wireless and human mobility battles the absence of global connectivity
When forwarding to any of a group of nodes, the power law coefficient increases significantly
All nodes are not equal
Some nodes are more active
Some pairs of nodes see each other more often
Difference in frequencies of connection opportunities within groups
Temporal influence on contact patterns

12. DTN Routing in a Mobility Pattern Space [2]
MobySpace
Formalism of DTN by a high-dimensional Euclidean space on node’s mobility patterns
Each dimension denotes the probability of finding a node in a specific location
Node connections arise and dissolve dynamically as a function of node mobility in physical space
Routing in DTN reduces to routing in the virtual space

13. Routing Concept
Each node’s mobility pattern is denoted by a MobyPoint in the MobySpace
Goal
Opportunistically forward a “bundle” (messages in DTN) to a node with a mobility pattern matching more and more to the destination
Action
Forward bundle to nodes with MobyPoint closer and closer to the MobyPoint of the destination

14. MobySpace Characterization
Contact based dimensions
Each axis is a possible contact with a node
The distance along that axis is the contact probability
Nodes with similar sets of contacts AND similar frequencies will be close in such a space
Location based dimensions
Each axis represents a specific location
The distance along that axis is the probability of finding the node in that location
Nodes visiting similar locations with similar frequency will be closer in such a space

15. Possible limitations and Issues
Not too effective if nodes change habits too often
Even under well defined mobility patterns, bundle may reach local maximum
In the location based case, 2 nodes may go to same locations with same frequency (MobyPoints coincide) but at different times

16. Case Study
Nodes move around N locations (dimensions) following a power-law distribution
P(i) is the probability for a node being at location i
P(i) = K * (1/d )^ni , where:
ni is the preference index of location i
d is the exponent of the power-law based mobility
K is a constant
K = (1 – 1/d )/(1 – 1/d N)
Higher d  smaller subset of preferred locations
Lower d  wider choice of locations for nodes

17. MobyPoint matching functions
Euclidean distance:
Canberra distance:
Cosine angle separation:
Matching distance:
Raw number of matching (within delta) location probabilities on an axis

18. Other methodologies Epidemic [6] Opportunistic Random
Whenever nodes meet they exchange bundles
Optimal path  minimum delay
High buffer occupancy and bandwidth utilization
Opportunistic
Always wait for destination only
One transmission per bundle
Random
When destination is not near, forward at random
Prevention of local loop added

19. Simulation Results – I (full knowledge)

20. Simulation Results – II (partial knowledge)

21. Practical Routing in Delay-Tolerant Networks [3]
Design Goals
Routing must be self-configuring
Devices deployed in remote regions
Must recover from failure without manual intervention
Acceptable performance over wide variety of connectivity patterns
Efficient use of buffer and network resources  Scalability factor
Network Model
Undirected graph with bidirectional links (contacts)
Unsuitable for unidirectional satellite networks
Contacts are assumed to have constant link bandwidth and delay

22. Based on previous work by others [7]
S. Jain, K. Fall, R. Patra, “Routing in a Delay Tolerant Network”, ACM SIGCOMM 2004
Knowledge Oracles
Contacts Summary Oracle
Time-invariant / aggregate characteristics of contacts
Contacts Oracle
Time-varying DTN multi-graph
Queuing Oracle
Buffer occupancy at any node at any time
Traffic Demand Oracle
Future traffic demand

23. Knowledge vs. Performance [7]

24. Partial Knowledge Routing [7]
Assigns cost to edges
Costs reflect estimated delay on edge
Queuing time: time till contact available
Transmission delay: time to inject into edge
Propagation delay: time to travel on edge
Computes minimum cost path
Conveniently uses different oracles
Computationally efficient distributed algorithms already exist for shortest-path based routing problems
Finds however single path from source to destination – no optimal splitting

25. Partial Knowledge Routing [7]
Cost function: ω (e, t)
If message arrives at node ‘u’ at time ‘t’, and if edge ‘e’ (between ‘u’ & ‘v’) is chosen, message will reach node ‘v’ at time ‘t + ω (e, t)’
FIFO: if t1 < t2, t1 + ω (e, t1) <= t2 + ω (e, t2)
Algorithm with Time-invariant Costs
Use modified Dijkstra’s algorithm
If L[v] > (L[u] + ω (e, L[u] + T)) then
L[v]  (L[u] + ω (e, L[u] + T)) (T  start time)
Minimum Expected Delay (MED)
Oracles: Contacts Summary
Edge cost = avg. wait time + prop delay + trx delay
Proactive routing is used for time-invariant cost
Fixed routes for all messages
Minimizes avg. waiting time but doesn’t optimize path
Improvement
Dynamically make use of superior contacts per-hop
Multiple disjoint paths to balance load

26. Minimum Estimated Expected Delay (MEED)
Does not require Contacts Oracle
Depends on observed mobility history
Aggregated mobility prediction over large time window
Time taken for message delivery in DTN
Nodes record connection and disconnection times over sliding history window
Tunable parameter for reaction time to changes
Topology distribution via epidemic link state routing

27. Who makes routing decision?
Source routing (NO)
In a DTN, source does not have end-to-end information
Per-hop routing (NO)
Intermediate nodes may not be able to route efficiently if topology changes
Per-contact routing (YES)
Routing tables are recomputed each time a contact arrives  most updated information
Pros
Is able to use contacts with high MED when present
Cons
Uses more resource for frequent re-computation
Adds computational delay to messages
May enter link loops (topology may change at each hop)

28. Topology distribution
Link state routing via Epidemic algorithm
Pros
Each node contains full topology
New node can get full information with one exchange
Cons
More memory required at each node  not scalable
Merging topology information amongst nodes is complex

29. Simulation Scenarios Mobility traces from Dartmouth WLAN  ad hoc DTN
More than 2000 users and 500 APs over 2 years
WLAN  ad hoc DTN
2 nodes are in contact if they are connected to same AP at the same time

30. Simulation parameters
30 nodes for one month
Nodes that at least contact another node 10 times in the month are included
Each node generates 6 messages (10,000 bytes) / 12 hrs
Protocols compared
Earliest Delivery (ED)  full contact schedule [7]
MED, MED Per Contact [7]
MEED
Epidemic [6]

31. Simulation Results - I

32. Simulation Results - II

33. Spray and Wait: An Efficient Routing Scheme for Intermittently Connected Mobile Networks [4]
Design Goals
Fewer transmissions per successful delivery
Low contention under high traffic loads
delivery delay close to optimal
Scalable w.r.t. network size or node density
Require low network knowledge

34. Spray and Wait - Concept
Spray phase
Every message is forwarded by source to L distinct “relays”
L is a number chosen to guarantee high delivery probability
Wait phase
Each of the L nodes wait for direct transmission only
Like SOLAR [11]
Nodes send/spray to subset of acquaintances only
Unlike SOLAR
Each acquaintance does both direct transmission and also relaying to further acquaintances
We choose the subset of acquaintances to guarantee high delivery probability

35. Spraying Techniques Source Spray and Wait Binary Spray and Wait
Spray to first L distinct nodes that come in contact
Binary Spray and Wait
Source of each message starts with L copies
At runtime, a node may have N messages (source + relay)
Upon contact with a node with NO copies (source or relay) a node with N messages hands over floor (N/2) and keeps ceil (N/2)
One copy left  direct transmission

36. Delay Comparison Theorem 1
When all nodes move in an IID manner, Binary Spray and Wait routing is optimal, that is, has the minimum expected delay among all spray and wait routing algorithms

37. Protocol Comparison Epidemic routing [6]
Randomized flooding with p = (0.02 – 1) [8]
Utility based routing with Uth = (0.005 – 0.2) [9]
Optimal Binary Spray and Wait with L copies
Seek and Focus single copy routing [10]
Oracle based Optimal Routing [7]

38. Simulation Results - I
100 nodes in 500 x 500 grid with reflective barriers
Random waypoint mobility model
Radio range is 10 grid units
Message inter-arrival time is Uniform (1, Tmax)
Tmax is varied from 10,000 to 2000
Average load of 200 to 1000 messages per unit time

39. Simulation Results - II

40. Simulation Results - III

41. IEEE Spectrum – August 2005 Issue

44. Extension of DTN in Space

45. References
Pocket Switched Networks and Human Mobility in Conference Environments - Pan Hui, Augustin Chaintreau, James Scott, Richard Gass, Jon Crowcroft, Christophe Diot, WDTN, ACM SIGCOMM 2005
DTN Routing in a Mobility Pattern Space - Jrmie Leguay, Timur Friedman, Vania Conan, WDTN, ACM SIGCOMM 2005
Practical Routing in Delay-Tolerant Networks - Evan P. C. Jones, Lily Li, Paul A. S. Ward, WDTN, ACM SIGCOMM 2005
Spray and Wait: An Efficient Routing Scheme for Intermittently Connected Mobile Networks - Thrasyvoulos Spyropoulos, Konstantinos Psounis, Cauligi S. Raghavendra, WDTN, ACM SIGCOMM 2005
The Interplanetary Internet - J. Jackson, IEEE Spectrum, Volume: 42  Issue: 8   Date:Aug. 2005, Pgs: 
Epidemic routing for partially connected ad hoc networks - Amin Vahdat, David Becker, Technical Report CS , Duke University, April 2000
Routing in a delay tolerant network - Sushant Jain, Kevin Fall, Rabin Patra, August 2004,  ACM SIGCOMM Computer Communication Review , Proceedings of the 2004 conference on Applications, technologies, architectures, and protocols for computer communications,  Volume 34 Issue 4
The broadcast storm problem in a mobile ad hoc network - Y.-C. Tseng, S.-Y. Ni, Y.-S. Chen, J.-P. Sheu., Wireless Networks, 8(2/3):153–167, 2002.
Probabilistic routing in intermittently connected networks - A. Lindgren, A. Doria, and O. Schelen. SIGMOBILE Mobile Computing and Communications Review, 7(3):19–20, 2003.
Single-copy routing in intermittently connected mobile networks - T. Spyropoulos, K. Psounis, and C. S. Raghavendra, In Proc. of IEEE Secon’04, 2004.
Sociological Orbit aware Location Approximation and Routing in MANET – Joy Ghosh, Sumesh J. Philip, Chunming Qiao, IEEE Broadnets 2005

46. Partial Knowledge Routing I [7]
Cost function: ω’ (e, t, m, s)
Edge ‘e’, time ‘t’, message size ‘m’, node assigning cost ‘s’
ω’ (e, t, m, s) = t’ (e, t, m, s) – t + d (e, t’)
where,
c (e, t)  capacity of edge ‘e’ at time ‘t’
Q (e, t, s)  queue size at source of edge ‘e’, at time ‘t’ as predicted by node ‘s’
t’  earliest time queued data at ‘e’ and message can be injected into the edge
Integral  volume of data through ‘e’ in interval [t, t’’]
d (e, t’)  propagation delay seen by message

47. Partial Knowledge Routing II [7]
Algorithms with Time-varying Costs
Earliest Delivery (ED)
Contacts Oracle
Q (e, t, s) = 0
Source routed
Buffer overflow  cascaded delay
Earliest Delivery with Local Queuing (EDLQ)
Q (e, t, s) = data queued for ‘e’ at ‘t’ if e = (s , *)
= 0 otherwise
Per-hop routed  path vector to avoid loops
Earliest Delivery with All Queues (EDAQ)
Contacts + Queuing Oracles
Q (e, t, s) = data queued for ‘e’ at ‘t’ at node s
Reservation of edge capacity along computed path