1. Analysis of a Campus-wide Wireless Network David Kotz Kobby Essien Dartmouth College September 2002

2. But how do people use it? How to design Wi-Fi networks? How to deploy Wi-Fi networks? Wi-Fi is becoming pervasive

3. Dartmouth College

4. 5,500 students 3,330 live on campus 40% own laptops

5. Class of 2006 just arrived 88% own Wi-Fi laptops; 100% of MBA students do

6. The Dartmouth network 476 access points (now over 500) Cisco model 340 or 350

7. Complete Wi-Fi coverage 200 acre campus 161 buildings –82 Residential –32 Academic –22 Administrative – 6 Library –19 Social Dartmouth campus

8. The largest Wi-Fi study Fall 2001 Wi-Fi at Dartmouth 11 weeks Over 1700 users Diverse population 161 buildings 476 access points Campus-wide coverage [Hutchins and Zegura] Wi-Fi at GA Tech 20 or 7 weeks 444 users Diverse population 18 buildings 109 access points Partial coverage [Balachandran 2002] Wi-Fi at SIGCOMM 2.5 days 195 users Computer scientists One room 4 access points Small sample [Tang and Baker 2000] Wi-Fi at Stanford 12 weeks 74 users Computer scientists 1 building 12 access points Small sample

9. Data Collection

10. 1: Syslog data collection AP reports interesting events –Authenticate –Associate –Deauthenticate –Disassociate Record date, time, MAC, AP name Sent by access point to syslog recorder

11. 2: SNMP data collection Every 5 minutes, poll each AP Record: –MAC of associated clients –Counter: inbound bytes –Counter: outbound bytes Inbound Outbound

12. 3: tcpdump data collection Tiny Linux box promiscuously sniffed all packet headers Hub AP

13. Results Traffic

14. Daily traffic (GB) Median: 53 GB/day

15. Average daily traffic, by weekday GB/day

16. Average hourly traffic, by hour To 11.4 GB (60% inbound) GB/hour

17. Median: 39 MB Average daily traffic per AP GB/day

18. Average daily traffic (GB) (by category) GB/day

19. GB/day/card Average GB/day per card (by category)

20. Average daily traffic (GB) (10 busiest buildings)

21. Results Sessions

22. time Associated Reassociated Deauthenticated Session duration 30 mins

23. Session length Hours Median session length is 16.6 minutes

24. Flickering sessions time Associated Reassociated 10 seconds

25. Max: 19,902 Median: 2 Roams per session Plot of the 18% of sessions that involve at least one roam

26. Results Users

27. Maximum: 77 days, 64 buildings, 161 APs Median: 28 days, 5 buildings, 9 APs Activity per card (distribution)

28. Active cards per day

29. Active cards per hour (category)

30. Results Protocols

31. Brown (dormitory)

32. Berry Library

33. Collis Student Center

34. Sudikoff Lab for Computer Science

35. Common IP protocols ProtocolGBPercent PIM~ 00.0% NARP~ 00.0% RSVP~ 00.0% IGMP~ 00.0% ICMP~ 00.0% UDP5.72.5% TCP221.997.5% Total227.595100.0% 99.7% of all wireless frames contained IP packets

36. TCP and UDP traffic (GB) (by building) These top ten account for 85% of traffic

37. TCP and UDP traffic (by direction)

38. TCP connections (by building)

39. TCP connections (by direction)

40. Correction: Figures 27-28 Dartmouth College Computer Science Technical Report TR2002-432 http://www.cs.dartmouth.edu/~campus/

41. Summary Largest trace-based study of a WLAN Large, diverse population –Residential university campus –Mixture of academic and residential patterns –Results may be different for other populations

42. Conclusions High variance in traffic and activity –From day to day, hour to hour –From place to place, user to user No clear dominance of inbound or outbound –Varies by protocol and user Dormitories dominated traffic –Especially the Tuck School of Business We need: –Cards that avoid roaming across subnet boundary –Support for roaming across subnets (Mobile IP, etc) –Symmetric bandwidth –Full-campus coverage: critical to acceptance

43. Thank you

44. David Kotz Kobby Essien http://www.cs.dartmouth.edu/~campus/