Cabernet: Vehicular Content Delivery Using WiFi

1. Cabernet: Vehicular Content Delivery Using WiFi Jakob Eriksson, HariBalakrishnan, Samuel Madden MIT Computer Science and Artificial Intelligence Laboratory Mobicom’08, September 14-19, 2008, San Francisco, California, USA Speaker: 黎焯杰

2. Abstract • Cabernet is a system for delivering data to and from moving vehicles using open 802.11(WiFi) access points(APs) encountered opportunistically during travel. • Two new components for improving open WiFi data delivery to moving vehicles • QuickWiFi: a streamlined client-side process to establish end-to-end connectivity with much less time than general. • Cabernet Transport Protocol(CTP): a transport protocol that improves throughput over TCP. • Deployed Cabernet on a fleet of 10 taxis in the Boston area, the result of which makes Cabernet a viable system for a number of non-interactive applications.

3. Contents Introduction Related Work Experimental Setting Design Overview and Goals Establishing Connectivity(QuickWiFi) Cabernet Transport Protocol(CTP) Vehicular WiFi Rate Selection End-to-End Performance Conclusion

4. Introduction As cars move, their connectivity is both fleeting and intermittent. Additionally, the packet loss rates over the wireless channel are both high and vary over the duration of a single AP association. Cabernet is well-suited for applications that deliver messages and files to users in cars, as well as from devices and sensors on cars to Internet hosts. The primary goal of Cabernet is to develop techniques that allow moving cars to obtain high data transfer throughput.

5. Introduction • Stock implementations of wireless networking protocols are not well suited to vehicular applications for three reasons. • Current client implementations take too long to establish end-to-end connectivity. • Current end-to-end data transfer and congestion control protocols do not work well when connectivity is fleeting and intermittent. • Default wireless bit-rate selection algorithms are tuned to non-moving users. • Cabernet incorporates three techniques to mitigate these problems. • Use QuickWiFi to reduce the establishing time. • Use CTP to improve end-to-end throughput over lossy wireless links. • Adopt a static 11Mbit/s bit-rate to improve link rates.

6. Contents Introduction Related Work Experimental Setting Design Overview and Goals Establishing Connectivity(QuickWiFi) Cabernet Transport Protocol(CTP) Vehicular WiFi Rate Selection End-to-End Performance Conclusion

7. Related Work • Cellular networks • The fundamental capacity of Cabernet is much higher than current cellular networks. • To complement cellular data services, not to replace them. • Vehicular networks • Drive-Through Internet, Infosations project, UmassDieselNet, etc. • Hadaller, Mahajan, etc. • Previous work in the CarTel project • Delay tolerant networks • Assume that mobile node are just a single hop from well-connected, fixed infrastructure. • Not focus on routing, but maximizing the utility of the single-hop path.

8. Contents Introduction Related Work Experimental Setting Design Overview and Goals Establishing Connectivity(QuickWiFi) Cabernet Transport Protocol(CTP) Vehicular WiFi Rate Selection End-to-End Performance Conclusion

9. Experimental Setting Developed a mobile vehicular testbed, hosted in taxis in the Boston area. Consisted of 25 nodes, each a Soekris 4801 computer, equipped with a Ubiquity Networks SR2(Atheros chipset) 802.11b/g radio, a small 3dBi omniderictional antenna mounted inside the pasenger compartment, a GPS receiver. Allocated 10 of the nodes for use in the Cabernet experiments. Successfully transferred data using 26,000 open APs The results were derived from a variety of experiment with varying duration and overlap. These measurements were truly gathered “in the wild”.

10. Contents Introduction Related Work Experimental Setting Design Overview and Goals Establishing Connectivity(QuickWiFi) Cabernet Transport Protocol(CTP) Vehicular WiFi Rate Selection End-to-End Performance Conclusion

11. Design Overview and Goals Scan for and attempt to associate with open Aps. Attempt to establish end-to-end connectivity. Use QuickWiFi to connect as quickly as possible. Use CTP to deliver data. Communicate with legacy Internet.

12. Design Overview and Goals Encounter: the interval between when the first beacon was heard from and AP and the time at which the last packet was heard from the AP. Stock implementations of the 802.11 and Internet protocols typically require several seconds to establish a connection(e.g. 12.9 seconds). QuickWiFi • Establish Connectivity Quickly • 5529 encounters with 1396 unique Aps while the cars in motion. • Used a single WiFi radio on each car. • Scanned continuously for open Aps until one was found, then associated with the AP and maintained the connection until the AP went out of range.

13. Design Overview and Goals The beginning and end of an encounter experience high los rates, even the middle is prone to losses. • Aps are not under control. • CTP implements a lightweight probing protocol to independently detect congestive losses on the path to the AP. • Investigate wireless bit-rate selection in the Cabernet setting to improve throughput. Most of these losses are caused by a combination of marginal links and mobility. A congestion control mechanism such as TCP’s that treats all end-to-end packet losses as congestion is sub-optimal. • Handle Non-Congestion WiFi Losses • Spent the duration of 850 random individual encounters sending bursts of back-to-back packets to the AP, with link-layer retries enabled, at the 11 Mbit/s-rate. • MAC layer ACKs were used to verify transmission success or failure.

14. Design Overview and Goals Most current applications and protocols do not tolerate intermittency. The CTP provides prompt feedback whenever end-to-end connectivity disappears and reappears. Allow applications to begin transfers when connectivity appears and avoid having to implement complex techniques to determine if connectivity is available. • Managing Intermittent Connectivity • Median time between successful encounters was 32 seconds and the mean was 126 seconds. • With 90% probability, a successful encounter will happen within 5 minutes after the end of the last encounter.

15. Contents Introduction Related Work Experimental Setting Design Overview and Goals Establishing Connectivity(QuickWiFi) Cabernet Transport Protocol(CTP) Vehicular WiFi Rate Selection End-to-End Performance Conclusion

16. Establishing Connectivity • Problems with Stock Implementations • Suboptimal scanning. • Inappropriate timeouts. • Sequential, yet asynchronous. • Manual intervention required. • Always-on connection model. Using stock implementations of various protocols can take significant time to complete, which is exacerbated in vehicular environment.

17. Establishing Connectivity • QuickWiFi Operation • Incorporates fully automatic scanning, AP selection, association, DHCP negotiation,, address resolution, verification of end-to-end connectivity, detection of the loss of connectivity. • Is implemented as a single state machine, running in one process, resulting a tight integration between steps: each is executed immediately after the previous has completed, and parallelism is exploited whenever possible. • Attempts to associate with the first open AP it encounters as it scans through the wireless channels; and resumes scanning at the point where it left off after a connection is complete. • In practice, this policy results in a quick way to select a random AP, since any AP beacon may be the first one heard when a scan begins.

18. Establishing Connectivity • Tuned for vehicular WiFi • Authentication, association, DHCP, and ARP all include timeout/retry protocols. • The mean connection establishment time can be dramatically reduced by reducing timeouts from seconds to hundreds of milliseconds. • Back-to-back authentication/association • No need to wait for the response to authentication request before sending the association request. • Two additional functions to monitor and report the status. • Ping-through--QuickWiFi uses a quick request-response exchange with a central server to verify end-to-end connectivity. • Connection loss monitoring--If any transmissions can not be seen for 500 milliseconds, it is likely that the car has moved out of the range of the AP.

19. Establishing Connectivity Let N be the number of channels, be the probability that an arbitrary AP is assigned to channel k. Let the strategy scans channel k at a rate . Using a fixed scanning strategy, the expected number of scans before the car successfully finds a single point on a random channel is proportional to . • Optimal Scanning Strategy • Passive scanning V.S. Active scanning • Only three channels in 802.11b/g are considered “orthogonal”: 1, 6, 11. • 83% of Aps observed are assigned to them, and 38.5% for channel 6.

20. Establishing Connectivity • Optimal Scanning Strategy • To find an assignment to that will minimize the number of scans, subject to . • The quantity is minimized when . • Using a Lagrange multiplier , define • Taking the partial derivative with respect to and setting to 0 yields • And thus , and . • Then generates a feasible scanning schedule, giving a weight to each channel k equal to .

21. Establishing Connectivity • QuickWiFi Performance • For each successful connection, measure the time between when the car discovered an open AP, and when QuickWiFi reported a successful connection to the application • Open AP Co-occurrence and AP Selection • QuickWiFi immediately initiates an association attempt as soon as an open AP is observed.

22. Establishing Connectivity • DHCP dominates the delay incurred, this stage takes a long time to complete. • DHCP servers send an ARP request prior to responding to a Discovery message. • ARP queries from the client spend less time than DHCP. • QuickWiFi Performance • Dissecting QuickWiFi latency

23. Establishing Connectivity • The vast majority of failures happen in the DHCP phases • Poor channel quality. • No MAC layer retries, with a corresponding increase in packet losses. • QuickWiFi Performance • Connection attempts fail

24. Contents Introduction Related Work Experimental Setting Design Overview and Goals Establishing Connectivity(QuickWiFi) Cabernet Transport Protocol(CTP) Vehicular WiFi Rate Selection End-to-End Performance Conclusion

25. Cabernet Transport Protocol • Cabernet uses a proxy to mediate between unmodified Internet hosts(e.g., servers) and the node in the car. • TCP performance was disappointing. • TCP with Explicit Loss Notification(ELN) • TCP Westwood/Westwood+ • CTP is an end-to-end, reliable, stream-oriented transport protocol that aims to achieve reasonable throughout for data transfers over short and loss-prone wireless connections. • CTP incorporates a different congestion control mechanism from the many previously proposed TCP variants to perform better over networks with high non-congestive loss rates.

26. Cabernet Transport Protocol • Reliability and Mobility • A CTP session does not break when the underlying IP address changes or path disappears. • CTP end hosts carry unique network-independent identifier, allowing CTP sessions to migrate seamlessly across changing IP address and Aps. • On an address change, the client reconnects to the other end and securely identies itself using its unique identifier. • An outgoing message is divided into packet-sized chunks, each of which is assigned a chunk sequence number. • ACKs are aggregated and sent at a moderate rate. • Each ACK contains a bitmap indicating what chunks are still missing. • The CTP sender retransmits chunks if they show up as missing more than a threshold number of times in the bitmap or after a timeout.

27. Cabernet Transport Protocol • CTP congestion control • CTP improves throughput by reacting only to congestive losses, which are assumed to occur on the path from the Internet sender to the AP. • Concurrent transmissions are handled by the 802.11 MAC prorocol running on the AP. • For transfers from a car to an Internet host, the CTP sender receives immediate feed back from the absence of link-layer ACKs, allowing the sender to retransmit data without reducing the transmission rate. • The challenge is to detect congestion on the wired part of the path without getting confused by wireless losses. • The CTP sender transmits probe packets periodically to the AP in question; a lack of response is interpreted as a sign of congestion.

28. Cabernet Transport Protocol • Probing Strategy • Three methods used by CTP, in order of preference. • TCP RST • ICMP TIMXCEED • ICMP ECHO • TCP RST is preferred because Aps that do provide a TCP RST response tend to do so without rate limitation. • ICMP probes tend to be rate limited to 1-2 per second. • Most Aps do not provide TCP RST response. • CTP relies on observed end-to-end packet losses for rate control, and adjust the rate on each end-to-end ACK.

29. Cabernet Transport Protocol • Rate adjustment • With feedback provided by end-to-end ACKs and periodic probe packets, CTP continuously adjusts the transmission rate for the connection. • Rate increase rule • The CTP sender increases its rate each time it gets an ACK. • Rate decrease rule • CTP reduces its rate every time a new probe packet is sent • Decrease rule has to reconcile some real world problems. • A packet loss probability , and a constant factor c in order to allow TCP sessions to successfully contend for link capacity.

30. Cabernet Transport Protocol • CTP vs TCP Performance • The measured median throughput achieved by CTP is approximately twice that of TCP. • The mean CTP throughput of 760 kbit/s was somewhat less than twice the TCP throughput of 408 kbit/s.

31. Cabernet Transport Protocol • CTP vs TCP Performance • The median CTP download is approximately twice the median TCP download. • The mean CTP download of 2438 kbytes is only 35% higher than the mean TCP download of 1860 kbytes.

32. Cabernet Transport Protocol • CTP vs TCP Performance • Long-duration encounters have significantly lower packet loss rate than short-duration encounters. • On long duration encounters, TCP is able to achieve nearly the same throughput as CTP.

33. Cabernet Transport Protocol • The measured TCP flow achieves higher throughput when competing with CTP than when competing with another TCP. • The standard deviation is lower when competing with CTP than with another TCP. • Interacting with TCP • Measured the downstream throughput of one TCP flow under three different conditions.

34. Contents Introduction Related Work Experimental Setting Design Overview and Goals Establishing Connectivity(QuickWiFi) Cabernet Transport Protocol(CTP) Vehicular WiFi Rate Selection End-to-End Performance Conclusion

35. Vehicular WiFi Rate Selection • 802.11b transmissions were much more likely to succeed than the faster 802.11g rates. • The difference in loss rate between 802.11b bit rates was small. • Even after losing packets, reducing the rate would not result in a performance. • Implemented a “fixed 11Mbit/s” rate selection strategy. The AP controls the download rate, so select rate for upload. Measured the loss rate of uploads from moving vehicles.

36. Contents Introduction Related Work Experimental Setting Design Overview and Goals Establishing Connectivity(QuickWiFi) Cabernet Transport Protocol(CTP) Vehicular WiFi Rate Selection End-to-End Performance Conclusion

37. End-to-End Performance • While Cabernet is not suitable for fast-paced browsing session, moderately interactive usage is feasible. • Cabernet is well suited to applications which perform periodic background updates as often as every few minutes. • Expected Transaction Time • The median response time for a small request was 2 min, mean 5 min. • The median response time for a large request was 4 min, mean 9 min.

38. End-to-End Performance • Expected long-term Throughput • Reduced TCP’s total download time of each encounter to account for the improvement in connection establishment time due to QuickWiFi. • The mean throughput for Cabernet was 38 Mbypte/hour(86 kbit/s).

39. Contents Introduction Related Work Experimental Setting Design Overview and Goals Establishing Connectivity(QuickWiFi) Cabernet Transport Protocol(CTP) Vehicular WiFi Rate Selection End-to-End Performance Conclusion

40. Conclusion • Vehicular network connectivity is intermittent, including high connection establishment latency and high WiFi lost rates. • Cabernet is designed to address those issues. • QuickWiFi, an optimized and integrated collection of tools for establishing connections with wireless access points, is used to reduce connection establishment time. • CTP is a transport protocol that handles high non-congestive wireless loss rates by running a lightweight probing protocol between a sender and the access point. • Cabernet is able to achieve ample throughput for a large class of non-interactive vehicular applications. • Future work: Open APs(for access point owners to “open up” and opt-in), Privately operated Aps.

41. Cabernet: Vehicular Content Delivery Using WiFi Thank You Very Much !

42. Effect of Antenna Type and Placement • An external antenna substantially improved signal quality. • The primary consequence of this improved signal quality was that the number of potentially usable Aps increased by 40%, accompanied by a corresponding decrease in the time between open AP encounters. The mean time between encounters was 35 seconds for the external antenna configuration, and 44 seconds for the internal antenna. Median time between encounters were 18 and 26 seconds respectively.