1. Probing in 802.11 using Neighbor Graphs Minho Shin, Arunesh Mishra, William Arbaugh University of Maryland

2. OUTLINE Motivation & Background Neighbor Graph NG / NG-prune probing algorithm Experiment/simulation results Conclusion

3. Motivation Hand-off latency is critical in WLAN infra. small coverage demand for multimedia or realtime applications High hand-off latency is observed 60 ~ 400 ms, 252 ms on avg, in experiment expected to increase with 11i authentications Probing latency > 90% of hand-off latency Reducing probing latency is important

4. Hand-off procedure AP1 AP2 AP5 MSAP1 [6] [11] [1] Reassociation Authentication AP5 AP4 AP3 AP2 [11] Probing AP3 AP4 [1] [2] [6] probe request probe response auth request auth response reassoc req reassoc res

5. Hand-off Decision Handoff Threshold AP1AP2 Hand-off Probing overhead Probing start hysteresis SNR from AP1 SNR from AP2 SNR ABCD Bad Signal High Threshold : - Too early probing, so high probing overhead Low Threshold : - Too late probing, so suffer from bad signal High Hysteresis : - Too late hand-off, so high probing overhead Low Hysteresis : - Too early hand-off, so ping-pong effect

6. What affects probing latency Number of channels to probe Standard doesn’t define naive : all 11 channels (Full-scanning) only used channels (Observed-scanning) Waiting Time for probe response Standard defines MinChannelTime / MaxChannelTime

7. Probing in detail ch1ch6 AP3AP5 MinChannelTime MaxChannelTime CS&T Wasted Channel Time Wasted Probe-wait MinChannelTime when no AP respond MaxChannelTime when any AP respond

8. Goals Do not probe empty channel prior knowledge of list of channels used by neighbor APs Do not wait more than needed prior knowledge of the list of neighbor APs use Neighbor Graph for prior knowledge [Mishra, Shin, Arbaugh, INFOCOM 2004]

9. OUTLINE Motivation & Background Neighbor Graph NG / NG-prune probing algorithm Experiment/simulation results Conclusion

10. Neighbor Graph Definition [Mishra, Shin, Arbaugh, INFOCOM 2004] NG dynamically learns the mobility patterns NG =, a directed graph V : set of all APs (AP i, AP j ) ∈ E if an STA handoffs from AP i to AP j Distributed data structure AP maintains the list of neighbor APs

11. Mobility Graph Given WLAN network and users C during period t Personal Mobility Graph PMG(c,t) : hand-off traces of client c during period t Mobility Graph MG is the aggregation of personal mobility graphs MG(C,t) = U PMG(c,t) c ∈ C MG =, a directed graph V : set of all APs (AP i, AP j ) ∈ E if an STA did handoff from AP i to AP j Neighbor Graph is an approximation of Mobility Graph

12. Neighbor Graph Generation H I D A E F J C B G A Personal Neighbor Graph (PNG) B Neighbor Graph (NG)

13. Neighbor Graph Dynamic Mobility Pattern edge-deletion deprecated path AP failure/removal H I D A E F J C B G A B edge-addition AP restore/install new path K

14. Quality of NG Edge addition hand-off costs high : Error Er(NG,t) = the ratio of # of edge-addtion to # of total hand-offs during t Storage overhead for deleted edge : Overhead Ov(NG,t) = ratio of # of edge-deletion * [time out] to ∑ of time of each edge’s residence Edge degree <= 6 in proper, non-redundant AP deployments

15. OUTLINE Motivation & Background Neighbor Graph NG / NG-prune probing algorithm Experiment/simulation results Conclusion

16. NG probing Key ideas Probe only neighbor channels Wait for only neighbor AP’s response ch1ch6 AP3AP5 MinChannelTime MaxChannelTime CS&T Saved Channel TimeSaved Probe-wait Next Channel

17. NG-prune technique AP1 AP2 AP5 [6] [11] [1] AP3 AP4 By NG probing, STA waits for MaxChannelTime AP2 and AP4 don’t overlap If STA knows Ap2 and AP4 doesn’t overlap, STA no longer waits for response from AP2 as soon as AP4 responds Use non-overlap graph

18. NG-prune Example F A E B C D [6] [11] [1] [6] [11] Channel 1 : 1*RTT Channel 6 : 1*RTT Channel 11: 0 Reduce both probe-count and probe-wait time Pruning algorithm can be reduced to a variant of the set cover problem [1]

19. NG-prune Comparison F A E B C D [6] [11] [1] [6] [11] Full Scanning : 2*MaxCT+9*MinCT Observed Scanning : 2*MaxCT+1*MinCT NG probing 1*Max + 1*Min + 1*RTT NG-prune probing 2*RTT [1]

20. NG-prune Non-overlap Graph Overlap Graph(OG) is an undirected graph ; V = set of access points ∈ E if their coverages overlap APi, APj overlap ⇔ S i (x) ≥ T h ∧ S j (x) ≥ T h Non-overlap Graph(NOG) = OG c OG is easier to generate than NOG Also distributed structure stored at each AP

21. OUTLINE Motivation & Background Neighbor Graph NG / NG-prune probing algorithm Experiment/simulation results Conclusion

22. What to measure Probing Latency (by experiments) Overal Probe channel count Probe wait time Performance (by simulations) vs number of channels vs number of neighbors

23. Experiment Methodology 20 Cisco 350 APs over two floors using channel 1, 6,11 Avg # of neighbors = 3.15 STA using laptop with Prism2 based wireless card Probing algorithm implemented in driver & user roaming program Full / observed scanning, NG / enhanced-NG

24. Experiment Probing Latency FullObsvdNGNGP Probe Count11.03.02.52.2 Probe Wait10.9ms11.0 ms6.3 ms4.4 ms NG-Pruning

25. Simulation Simulation Model Identical coverages Randomly chosen : # of neighbors positions STA’s direction Optimaly chosen channel assignments Variables # of Nb : 2,3,…,8 # of Chnl : 3,5,8,12

26. Simulation Probing Latency vs # of Channels

27. Simulation Pruning vs # of Neighbor

28. Conclusion New efficient probing algorithms ( NG/NG-prune probing ) and evaluated by experiemt and simulations Performance improves as # of indep. channels increase (802.11a) as density of access points increase (# of neighbor increase)

29. Thank you