1. 1 Self-Routing in Pervasive Computing Environments using Smart Messages Cristian Borcea, Chalermek Intanagonwiwat, Akhilesh Saxena (Rutgers University), and Liviu Iftode (University of Maryland)

2. 2 Networks of Embedded Systems Linux Cell PhoneLinux WatchLinux CameraLinux Car Functionally heterogeneous nodes Very large scale Ad hoc topologies Dynamic network configurations Limited a priori knowledge about network resources

3. 3 Programmability Challenge Traditional message passing distributed computing does not work for networks of embedded systems –unknown and volatile network configurations –end-to-end data transfer may hardly complete (i.e., “all or nothing” semantics is not appropriate) –fixed address naming and routing (e.g., IP) are too rigid Our Solution: Cooperative Computing using Smart Messages More flexible naming and routing are needed –applications interested in content/services, not individual nodes –different applications have different routing requirements

4. 4 Outline Motivation Smart Messages Overview Self-Routing Mechanism –content-based migration –application scenarios Evaluation –SM prototype –simulations Related Work Conclusions & Future Work

5. 5 Smart Messages at a Glance Distributed computing using execution migration Applications composed of one or multiple Smart Messages Smart Message (SM): –composed of code, data, and execution state –executes on nodes of interest named by properties Cooperative Nodes: –execution environment (Virtual Machine) –content-based memory (Tag Space) Self-Routing –routing performed at application-level –applications can change routing during execution

6. 6 Application Example n=0 while (n<2) migrate(Taxi); if (readTag(Available)) writeTag(Available, false); writeTag(Location, myLocation); n++; n=0 Taxi n=0 n=1 n=2 need 2 taxis mobile data application routing

7. 7 Node Architecture Admission Manager Virtual Machine SM arrival SM migration sm1 sm2 …… tag1 tag2 …… Tag Space Admission prevents excessive use of resources Execution is non-preemptive, but time bounded Two types of tags: – application tags – I/O Tags Tags used by SMs for: – naming – storage – synchronization – I/O access SM Ready Queue

8. 8 Smart Messages Migration 1234 migrate(Taxi) sys_migrate(2)sys_migrate(3)sys_migrate(4) Two level migration: – migrate() embeds routing algorithm migrates application to next node of interest names nodes in terms of arbitrary conditions on tag names and tag values – sys_migrate() one hop migration used to implement migrate Taxi

9. 9 migrate(Taxi){ while(!readTag(Taxi)) if (readTag(RouteToTaxi)) sys_migrate(readTag(RouteToTaxi)); else create_SM(DiscoverySM, Taxi); createTag(RouteToTaxi, lifetime, null); block_SM(RouteToTaxi, timeout); } migrate(Taxi){ while(!readTag(Taxi)) if (readTag(RouteToTaxi)) sys_migrate(readTag(RouteToTaxi)); else create_SM(DiscoverySM, Taxi); createTag(RouteToTaxi, lifetime, null); block_SM(RouteToTaxi, timeout); } Migration Example 12i RouteToTaxi = 2 Taxi RouteToTaxi = ? Network migrate(Taxi){ while(!readTag(Taxi)) if (readTag(RouteToTaxi)) sys_migrate(readTag(RouteToTaxi)); else create_SM(DiscoverySM, Taxi); createTag(RouteToTaxi, lifetime, null); block_SM(RouteToTaxi, timeout); } RouteToTaxi = j

10. 10 Self-Routing Smart Messages carry the routing and execute it at each node Smart Messages control their routing –select routing algorithm (migrate primitive) from multiple library implementations implement a new one –change routing algorithm during execution in response to adverse network conditions according to application’s requirements

11. 11 Dynamic Change of Routing (1) while (n<3) try{ migrate(tag, timeout1); }catch(TimeoutException e){ migrate(tag, timeout2); } Dense network Low mobility Proactive routing Sparse network High mobility On-demand routing

12. 12 Dynamic Change of Routing (2) migrate(circle); while (n<3) migrate(tag); geographical routing to reach circle space-bound on-demand routing to reach the nodes of interest

13. 13 Smart Messages Routing Algorithms Goal: Evaluate the potential of SMs to implement different content-based routing algorithms –on-demand content-based routing (similar to AODV [Perkins ’99]) –greedy geographical routing (similar to GPSR [Karp ’00]) –proactive routing using Bloom filters (similar to Probabilistic Routing [Rhea ’02]) –rendez-vous routing (combining on-demand and proactive routing) e.g., geographic dissemination + limited flooding advantage: improves the response time for applications while avoiding global dissemination and large scale flooding

14. 14 Evaluation Strategies Implementation –SM prototype over Sun Java KVM on HP iPAQs –small scale network (8 nodes) –evaluated the effects of code caching Simulation –SM simulator –large scale network (256 nodes) –evaluated the effects of best routing selection and dynamic change of routing

15. 15 Prototype Implementation Modified version of Sun Java KVM HP iPAQs running Linux 802.11 for communication Completion Time Routing algorithmCode not cached (ms)Code cached (ms) Geographic On-demand 415.6126.6 506.6314.7 user node node of interest intermediate node

16. 16 Simulation Event-driven simulator extended with support for SM execution Setup: –256 nodes uniformly distributed over a 1000m by 1000m square –transmission range = 100m –bandwidth = 2Mbs –each node has an average of 6 neighbors (min = 2, max = 11) Metrics: –completion time: user-observed response time for an application –total number of bytes sent: total amount of traffic generated by an application also indicates the energy and bandwidth consumed by an application

17. 17 Simple On-Demand Routing vs. Conditional On-Demand Routing starting nodenode of interestother node nodes of interest contain a certain tag, and the tag’s data must satisfy a given condition 5 nodes (distributed uniformly) have the given tag vary the number nodes of interest from 1 to 4

18. 18 On-demand Routing vs. Geographical + On-demand Routing starting nodenode of interestother node 5 nodes of interest distributed over the red region – radius = 500m application has knowledge about the desired region – vary the radius from 500m to 1500m

19. 19 On-Demand Routing vs. Geographical + On-Demand Routing Cont’d starting nodenode of interestother node 3 nodes of interest located in the corners – have to be visited in clockwise order application has knowledge about these nodes’ regions – vary the radius from 100m to 700m

20. 20 Related Work Mobile agents –e.g., D’Agents[Gray ‘97], Ajanta[Karnik ‘98] Active networks –e.g., ANTS[Wetheral ‘99], SNAP[Moore ’01] Mobile ad hoc networking –e.g., DSR[Johnson ‘96], AODV[Perkins ‘99], GPSR[Karp ‘00] Content-based naming and routing –e.g., INS[Adjie-Winoto ‘00], CBR[Gritter ‘01] Pervasive computing models –e.g., one.world[Grimm ‘01] Sensor networks –e.g., Diffusion[Intanagonwiwat ‘00], TinyOS[Hill ‘00]

21. 21 Conclusions Self-Routing provides high flexibility for SM applications –choose the routing –implement their own routing –change the routing dynamically Self-Routing has performance benefits –improved response time for applications –significant energy and bandwidth savings in the network

22. 22 Future Work Spatial Programming with Smart Messages –programming model for networks of embedded systems –network resources accessed transparently using {space, tag} spatial references –a node referenced by {space, tag} is reached through a combination of geographical and content-based routing Design and implement real world applications using Smart Messages and self-routing

23. 23 Thank you! http://discolab.rutgers.edu/sm