1. 1 A Cross-Layer Architecture to Exploit Multi-Channel Diversity Jay A. Patel, Haiyun Luo, and Indranil Gupta Department of Computer Science University of Illinois at Urbana-Champaign Distributed Protocols Research Grouphttp://kepler.cs.uiuc.edu/

2. 2 Motivation: Mesh networks do not scale Wireless mesh networks: “Killer app” –MIT Roofnet –Champaign-Urbana Wireless Contention: single channel –Intra-flow interference –Inter-flow interference –Worsens near gateway(s) Gateway node Can a single “commodity” transceiver exploit multi-channel diversity?

3. 3 Challenges + Prior Work Neighbors must converge to exchange data –While exploiting multiple channels Locally opportunistic channel hopping –Multi-channel MAC [So:MobiHoc04] –Seeded Slotted Channel Hopping [Bahl:MobiCom04] Limitations –Leads to node synchronization problem –MAC Approach: Probable implementation issues

4. 4 Our Contributions Dominion: A cross-layer architecture –Simple MAC + Intelligent routing –Key decisions shifted up, i.e., in to the software stack Deterministic channel hopping MAC protocol –Eliminate locally opportunistic behaviour Improves fairness Core logic resides at the routing layer –Graph-theoretic model: extensible and flexible –Multi-path routing

5. 5 Split Topology: k subnetworks Frequency Division + CSMA Approach –Logical subnetworks: A subnetwork per channel –Node n i homed at channel SHA1 (n i ) mod k –Creates network and subnetwork partitions Route across network partitions? f1f1 f2f2 f3f3

6. 6 Time is on our side... Key: Periodically converge subnetworks –Each pair of subnetworks switches to a common channel at a pre-determined time “Deterministic scheduling” –Based on modulo arithmetic –Can be generated simply with the parameter k –MAC uses this schedule Primary difference vs. IEEE 802.11

7. 7 A Sample Schedule s2s2 s3s3 s4s4 s5s5 s0s0 s5s5 s2s2 s3s3 s4s4 s4s4 s0s0 s1s1 s5s5 s3s3 s5s5 s4s4 s0s0 s1s1 s2s2 s2s2 s3s3 s5s5 s0s0 s1s1 s3s3 s1s1 s4s4 s2s2 s0s0 s0s0 s1s1 s2s2 s3s3 s4s4 s5s5 t0t0 t1t1 t2t2 t3t3 t4t4 k = 3 f2f2 f3f3 f1f1 Number of subnetworks: 2k Schedule cycle: T= NextPrime(2k - 1) Exactly 2 subnets converge on a channel Every subnet converges every other subnet s1s1

8. 8 Connectivity : A Visual Guide DominionIEEE 802.11

9. 9 Routing Best route for A -> B? –Two routes: AB (direct) and AC -> CB (indirect) Which is the better route? It depends –Throughput-wise: AB Can we do better? YES! with multi-path routing –Latency-wise: is time-variant Addressed in a follow-up paper A [s 2 ] B [s 3 ] C [s 0 ] t4t4 t2t2 t1t1

10. 10 Abstraction: Graph-Theoretic Model Convert link state to an abstract model Edge weight assignment –Connectivity edge = p f, temporal edge = 0 Locate shortest route using Dijkstra’s Multi-path routing –Prune all connectivity edges in route –Repeat: until no more routes found A5A5 A0A0 A1A1 A2A2 A3A3 A4A4 C1C1 B4B4 Temporal Edge Connectivity Edge Base Edge A [s 2 ] B [s 3 ] C [s 0 ] t4t4 t2t2 t1t1

11. 11 Experiment Methodology Implementation –QualNet v3.9 –10 ms timeslots, 80 µs switching delay Only 11 channels used (out of 12 for 802.11a) Topology –100 nodes, 1000m x 1000m –Uniform random placement –Random assignment of nodes to subnetworks

12. 12 Results Distance-normalized aggregate throughput: Dominion vastly better than SSCH (86%) and 802.11 (1813%)

13. 13 Results (continued) Jain’s fairness index shows that Dominion is fair –1730% fairer than 802.11, and 315% fairer than SSCH

14. 14 Conclusion New cross-layer architecture –Dominion exploits k channels with only 1 radio –Eliminate locally opportunistic behavior Simple MAC: deterministic schedule –Intelligence shifted upwards Suitable for static, wireless mesh networks –Excels in non-disjoint multi-flow scenarios Distributed Protocols Research Grouphttp://kepler.cs.uiuc.edu/

15. 15 Questions

16. 16 Future Work Dynamic subnetwork assignment –Based on two-hop “neighborhood” Extend the Graph-theoretic model –Optimize on end-to-end latency TCP improvement –Multiple routes leads to out-of-order packets Broadcast packets –Probabilistic approach –Allow efficient dissemination of link-state at run-time

17. 17 Implementation QualNet v3.9 10 ms timeslots, 80 µs switching delay Source routing Per-flow, per-timeslot queuing –prevents head-of-line blocking Warnings reduce buffer overflow at intermediate nodes Attempts only 1 DCF transmission per packet at a time –Allows for on-time switching A packet is dropped after 14 DCF failures –akin to two 802.11 retries

18. 18 Experiment Methodology Implementation –QualNet v3.9 –10 ms timeslots, 80 µs switching delay 100 nodes, 1000m x 1000m –Uniform random placement –Random assignment of nodes to subnetworks Bootstrap process: measure quality of each link –802.11 and SSCH: used to calculate static ETX routes –Dominion: network link-state Results are average of 5 independent trials –Only 11 channels used (out of 12 for 802.11a)

19. 19 Multi-Path Routing Using Dijkstra, locate shortest route Prune all connectivity edges in route –Reduces or eliminates inter-flow interference Repeat: until no more routes found

20. 20 Outline Motivation Related Work Dominion: Key Contributions Deterministic Scheduling Routing Intelligence Experimental Results Conclusion