Vehicle-to-Infrastructure (V2I) Traces
We have collected traces of wireless network quality for two cellular networks (Sprint and Verizon) and one commercial WiFi hotspot (XFinityWiFi) that has been deployed rapidly in densely-populated area. The traces are collected in four representative driving scenarios - downtown, highway, rural, and suburban areas - to measure wireless network quality in each of unique scenarios. The traces were collected in October 2017 around Ann Arbor, each ranging from 40-62 minutes length.
For trace collection, we configured VNperf to measure every 1 second interval to measure. We also configured VNperf to declare that latencies of over 5 seconds to represent blackout (unavailable) periods. We found that for WiFi hotspot, this is a common case because of low signal. We also found out that XFinityWiFi supports seamless WiFi-roaming that queues packets from one AP when device is disconnected and deliver to another AP when reconnected. Queued packets from one AP can later deliver over another AP at reconnection. To avoid previous queued packets affects latency of present measurements, VNperf logs -1 to declare the network is unavailable or exhibits high latency, if more than 5 previous measurements have not been successfully completed. Thus, periods of network unavailability appear to be intervals that exhibit extremely high latency.
For WPA authentication, XFinityWiFi driver handles WiFi-roaming based on network interface's MAC address. When more than one access points (APs) are available, we configured WPA supplicant to associate the one with the highest signal strength. We noticed that DHCP sometimes incorrectly triggers when the interface has not been associated with any AP for more than 5 minutes. So, we used DHCP enter-hook to detect new gateways at every WPA association.
We collected traces in October 2017, each ranging from 40-62 minutes in length. Trace D1 was collected driving through the downtown areas of Ann Arbor, MI (population approximately 120,000 and metro area population approximately 350,000). Trace D2 was collected solely on interstate highway driving, primarily but not exclusively through rural areas. Trace D3 was collected on rural roads in sparsely-populated areas. Trace D4 was collected in suburban locations that included neighborhoods, subdivisions, and secondary roads.
Summary of traces
The following is a shortened summary of trace analysis. Details of study is available in Table 1 of our paper, RAVEN: Improving Interactive Latency for the Connected Car.
Throughout the four traces, we found that
- No network consistently offers the lowest RTTs.
- It appears quite challenging to predict which network will lower the lowest RTT over short time scales.
- At the tail of each CDF, RTTs are very high, which will substantially degrade interactive applications.
- High RTTs are weakly correlated. High RTTs on one network could be masked by using another network.
For each row, there are 6 columns:
|vehiclespeed||Speed of vehicle|
|latitude||Latitude of vehicle|
|longitude||Longitude of vehicle|
|verizon||RTT over Verizon|
|sprint||RTT over Sprint|
|xfinitywifi||RTT over XFinityWiFi|
Each row contains measurements of Verizon, Sprint, and XFinityWiFi along with vehicle speed and location. As noted previously, we configured VNperf to measure at every 1 second interval to minimize external variables. For each scenario, the vehicle was driven without previously designed route to remove bias. However, the speed limit in each route was strictly enforced.
Visual Route Map
Each map.d[1,2,3,4].html contains graphical overview of routes taken in each scenario around Ann Arbor. To view the map,
$ cd vnperf/traces $ git clone https://gist.github.com/4dec208e28b36ce554d43c8653123050.git maps/ $ mv maps/*.html . $ python -m SimpleHTTPServer # open your favorite web browser # goto http://localhost:8080
Replay Traces in a Testbed
Replay script emulates latency changing condition by replaying a selected trace in testbed.
# make sure you have Linux TC setup $ cd vnperf/traces $ python replay/replay.py public.d1_downtown.csv
All source code, documentation, and related artifacts associated with the cloudlet open source project are licensed under the Apache License, Version 2.0.