1. Disruption Tolerant Networks Aruna Balasubramanian University of Massachusetts Amherst 1

2. What? Termed coined by DARPA Fundamentally different way of looking at networks Internet Wired LAN Wireless LAN 2 Cell tower End to End connectivity between device and Internet gateway

3. Primary characteristics of DTNs: No contemporaneous end-to-end path need to exist 3

4. Traditional networks i Source Destination 4 Intermediate

5. Disruption Tolerant Networks Post office model Store and forward i XZY i i 5

6. Why bother? Can be adapted to scenarios other than inter- planetary communication To enable network access, when infrastructure is  difficult to deploy  expensive to deploy  available, but a DTN can still improve performance  Some real life examples where DTNs are being used….. 6

7. Infrastructure is difficult to deploy Wild-life tracking TurtleNet project, UMass Deployed in Amherst ZebraNet project, Princeton Deployed in Mpala, Kenya 7

8. Infrastructure expensive to deploy Providing Internet connectivity to developing regions. E.g., Email KioskNet in Waterloo, Digital Gangetic Project in India 8

9. Even when infrastructure is available Provide a cheaper alternate to cellular data plans. Google from the bus without a 3G plan!!! DieselNet project, UMassCarTel project, MIT 9

10. Outline Why are DTNs useful Application layer: How are the applications really implemented? Routing layer DTN stack Power management Lessons learnt from our deployments efforts 10

11. Single hop case When node meets an Internet gateway it sends/receives data. Internet 11

12. How to implement email in single hop? Email protocols today (IMAP/POP) cannot work in the presence of disruptions Solution: Use a gateway 12

13. How about web search? Retrieving web…. Retrieving images… Retrieving…. 13

14. Web search challenges 14 Frequent disruptions may mean you keep retyping the query

15. Adapting web search to mobile networks (Thedu) Queries from mobile Store query Interface Google, Yahoo, Live, Ask, …. Google, Yahoo, Live, Ask, …. Snippets Prefetch Store web pages Store web pages Web pages returned to mobile Thedu proxy Thedu Client 15

16. Other web search apps based on DTNs The TEK search engine: Collects all possible data for a search query and returns it in email format RuralCafe  Caches search queries to perform local search  Send search responses in the form of sessions 16

17. Outline Why are DTNs useful Application layer Routing layer: How can we support multihop? DTN stack Power management Lessons learnt from our deployments efforts 17

18. Routing challenges Wired/Mesh/MANETs End-to-end path exists Known topology Low feedback delay  Retries possible DTNs  No end-to-end path  Uncertain topology  Feedback delayed/nonexistent Primary challenge: finding a path to the destination under extreme uncertainty 18

19. Post office model may not always work! i X Z Y i 19

20. Key idea in DTN routing: Replication i XZY i i W i Naïve replication using flooding wastes resources and can hurt performance 20

21. Efficient replication When two nodes X and Y meet, what packets should be replicated? Heuristics  Random replication: X randomly select packets in the buffer and transfer to Y  Maximum replication count: Set a replication threshold for each packet  Meeting frequency: X will send a packet to Y, if Y has a higher probability of meeting the destination. 21

22. More replication-based heuristics Utility-based routing (Our work)  Each packet is given a utility, based on the routing metric.  For example, if the routing metric is to minimize delays, the utility is the expected delivery delay  The first packet replicated is one whose replication decreases the delivery delay by most 22

23. Outline Why are DTNs useful Application layer Routing layer DTN stack: Can we simply use the wireless stack? Power management Lessons learnt from our deployments efforts 23

24. Do we need a new link layer? 802.11 is successful link layer protocol for wireless networks. Average time to connect ~ 13 sec Average duration of AP meeting in DieselNet is 25 sec 24 In DTNs, the timeouts and retries and significantly reduced. DHCP overhead is reduced by caching

25. DTN2 stack The DTN stack has two additional components  A bundle protocol: For store and forward  A convergence layer: To determine sending rate Being standardized by IETF, implemented by BBN 25 Sender Receiver

26. Outline Why are DTNs useful Application layer Routing layer The DTN stack Power management Lessons learnt from our deployments efforts 26

27. Power Management Motivation: To have perpetual battery-operated network systems Example: If GPS is on, battery life is 3 hours 27

28. Key idea for power management: Energy Harvesting Use solar cells to scavenge energy Challenges Amount of energy harvested depends on size of the cell Variable energy harvested per node Seasonal, unpredictable Take away: Smart power management scheme needed even with energy harvesting 28

29. Outline Why are DTNs useful Routing layer challenges Link and transport layer challenges Application layer challenges Power management challenges Lessons learnt from our deployments efforts 29

30. UMass DieselNet 30

31. Details 40 buses, 26-node mesh testbed Our lab pays $1600 per month for 3G connection on buses; no monthly cost for WiFi Roughly 50GB of data is downloaded from the bus using WiFi 31

32. DieselNet Advantages Very useful for research: Evaluation is a lot more believable; forced to think practical Useful for the community. Example: bus tracking project, pothole patrol 32

33. Challenges in outdoor deployment Difficult to fix broken parts Cannot predict the quality of information collected, because  Many buses may be broken  Maybe running different versions Bomb scare!! 33

34. Take Aways DTNs useful in various environments Protocols that work well in wired and even wireless networks do not work well in DTNs Rethink all four layers of the OSI stack, as well as power management 34

35. Resources DTN research group: http://www.dtnrg.org/http://www.dtnrg.org/ DieselNet, TurtleNet: http://prisms.cs.umass.edu/dome/ http://prisms.cs.umass.edu/dome/ MIT’s CarTel: http://cartel.csail.mit.edu/http://cartel.csail.mit.edu/ Waterloo’s KiokNet: blizzard.cs.uwaterloo.ca/tetherless/index.php/Kio skNet: blizzard.cs.uwaterloo.ca/tetherless/index.php/Kio skNet My website: www.cs.umass.edu/~arunabwww.cs.umass.edu/~arunab 35

36. Challenge in deploying Email, FTP (2) TCP throughput very low in the mobile setting  Starts sending 1 packet per window  Increases packets by 1 per window if not losses  If a single packet is lost, the window size is halved.  TCP thinks losses are due to congestion, and another node is sending Even if the bandwidth is 1Mbps, TCP only uses a small portion of the bandwidth Possible solution: Make TCP differentiate between congestion and bad channel quality. Decrease rate only for congestion. Possible solution: Make TCP differentiate between congestion and bad channel quality. Decrease rate only for congestion. 36

37. Link and transport layer challenges X ZY Link layer challenges: similar to any other network, except in handling handoffs during mobility Transport layer challenges: TCP, UDP are end-to-end protocols. But there is no end-to-end connectivity OSI Stack 37