A Measurement Study of Vehicular Internet Access Using In Situ Wi-Fi Networks Vladimir Bychkovsky, Bret Hull, Allen Miu, Hari Balakrishnan, and Samuel Madden MIT CSAIL http://cartel.csail.mit.edu Wi-Fi Is Everywhere What are the performance properties of organically grown Wi-Fi networks? Images from WiGLE.net and CarTel • Today: The Opportunity – Broadband connections are often idle – 65% of on-line households have Wi-Fi • What if … – … home users open up their APs … – … and share/sell the spare bandwidth? • Cellular alternative for mobile users: – Messaging (multimedia, e-mail, text) – Location-aware services – Mobile sensor networks (e.g. MIT project CarTel ) • Challenges – Legal, economic, security, policy issues – Performance Wi-Fi For Mobile Messaging: Will it work? • Wi-Fi cells are smaller than cellular cells – Is density sufficient? Are connections too short? • Organically grown, unplanned deployments – Uneven densities, AP churn, unpredictable • Back-of-the-envelope: – 55 km/hour: ~15 meters/s – ~150 meter AP coverage [Akella’05] – ~10 sec connectivity • What about connection overhead? – scan, associate, get IP, etc. • Current stacks too slow – How long does it take your laptop to get an IP here? Outline • Data and experimental method • Connectivity properties • Data transfer properties • Towards OpenWiFi networks Deployment and Data • 232 days of normal driving (07/05 – 07/06) Area shown: ~21x15 km – in Boston and Seattle – 290 hours of clean data – 260 distinct km of roads • 50% data from 15 km – 32,000 APs discovered • 2000 open – 75,000 AP join attempts • 9 cars: – Embedded PC – 200mW 802.11b @ 1MBps – GPS unit GPS unit Wi-Fi Antenna Experimental Method: Scanping No access points found scan open AP found get ip IP in cache? use cache try DHCP success associate success 3 seconds of lost pings get ip local AP ping success e2e ping success tcp test upload ping success Association Duration Definition scan associate get ip IP acquisition time AP ping loss Fraction of successful attempts IP Address Acquisition Cached IP Combined DHCP • Default DHCP timeout is too long • Simple fixes: • small DHCP timeout • caching leased IP IP acquisition delay (s) Outline • Data and experimental method • Connectivity properties • Data transfer properties • Towards OpenWiFi networks Association Duration Definition 1st AP ping Last AP ping received received scan associate get ip AP ping association duration time loss Fraction of associations Association Duration Associations last over tens of seconds even at vehicular speeds. • Median: 13 seconds • Mean: 24 seconds Association duration (s) Fraction of associations Connectivity vs. Speed Connections established at range of speeds. Little data at higher speeds (system is not optimized for subsecond connections yet) Speed (km/h) Association Duration (s) Association Duration vs. Speed ~10 seconds at 55km/h Speed (km/h) Estimating AP Coverage Procedure: 1. Note locations 2. Find bounding box 3. Report diagonal 200 ft 100 m location at the time of connection Fraction of access points Access Point Coverage Open Wi-Fi access points have a significant coverage area even in urban setting. • Median: 100 m • Mean: 150 m Diameter of AP coverage (meters) Fraction of successful scans Urban Access Points Density Access points are highly clustered. Using multiple access points at the same time may further increase throughput. Number of APs discovered per scan Time To Connectivity Definitions End-To-End connection Join Success (no e2e) Join failed (MAC filtering) Fraction of events Time Between Connectivity During normal driving we encounter a new access point every 23 seconds on average. Today we can only use one every 260 seconds on average. Join Attempts Join Successes E2E Success Time between events (s) Outline • Data and experimental method • Connectivity properties • Data transfer properties • Towards OpenWiFi networks Fraction of connections Bytes Uploaded Per Connection Non-trivial amount of data: Median: 200 KBytes per connection Mean: 600 KBytes Consistency check: 600 KBytes / 24 sec = 25 KBps Bytes received on server (KBytes) Packet delivery rate Impact of Mobility on Delivery Rate 80% delivery rate would cripple TCP Hypothesis: losses are non-uniform Speed (km/h) Related Work • Location and range of in situ Wi-Fi: – wardriving.com, wigle.com, wifimaps.com – Akella et al ’05, ‘06 • Vehicular Mobility of Wi-Fi client: – Ott and Kutscher ’04, ’05; Gass et al ’06; etc • Mobility in cellular networks: – Rodriguez ’04; Qureshi and Guttag ’05; etc This is the first end-to-end Wi-Fi performance study under normal driving conditions Outline • Data and experimental method • Connectivity properties • Data transfer properties • Towards OpenWiFi networks Towards Open Wi-Fi Networks • Today – Rampant, high-bandwidth use is a bad idea • “Unauthorized access” or “trespassing” • May violate ISP contract even if users “opt-in” • Solution: – Part I: provide economic incentives (Fon, etc) • Mobile user pay nominal fee • Home users “opt-in” • ISPs get a cut – Part II: provide technology • Tiered accounting, security, and QoS for home APs • Fast delay-tolerant stack for mobile users Conclusion • Today, during normal driving – – – – – New access point every 23 seconds (avg) Associations last for 24 seconds (avg) Median TCP upload: ~200 Kbytes Connectivity is equi-probable at [0; 60] km/h In situ APs are is highly clustered • Use multiple APs simultaneously – Simple techniques can improve DHCP latency OpenWiFi networks have tremendous potential. Will we tap into it? http://cartel.csail.mit.edu