Overview
---
The accompanying data files represent bandwidth and latency measurements
gathered for the evaluation of the system described in "Intentional
Networking: Opportunistic Exploitation of Mobile Network Diversity," by Brett
D. Higgins, Azarias Reda, Timur Alperovich, Jason Flinn, T.J. Giuli, Brian
Noble, and David Watson, in Proceedings of the Sixteenth International
Conference on Mobile Computing and Networking (MobiCom 2010).  Please refer
to this paper for details on how these traces were used.

Please direct any questions, comments, etc. to Brett Higgins
(brettdh@umich.edu).


Methodology
---
To gather these traces, we used a simple measurement application, similar to
Virgil [2], that connects to a reference server in our on-campus lab and
performs a sequence of tests: downstream bandwidth, upstream bandwidth, and
round-trip time. Our mobile testing computer was a laptop with two network
interfaces: WiFi and an ethernet-connected Sprint 3G device. The bandwidth
recorded is based on TCP goodput. Each test ran for at least 1 second; we ran
down/up/rtt test groups continuously with 1-second gaps in between groups,
recording the end-of-test timestamp along with the results.

We only record a WiFi encounter in the trace if we are able to associate to
the access point, obtain an IP address, and connect to our reference server.
Commonly available DHCP implementations have long timeout periods - 1 second
or longer. Since packet loss is common in wireless networks, DHCP lease
acquisition can often take a few seconds. We need to expedite this process,
as many vehicular WiFi encounters only last a few seconds. Thus, we
implemented a DHCP client similar to the design of QuickWiFi [1]. By reducing
the timeout value and aggressively sending more requests, we achieve lease
acquisition times on the order of 500 ms, allowing us to measure more of the
available WiFi APs.

Tests were run independently on each interface, which means that measurements
for one interface are not perfectly synchronized with measurements on the
other interface.  We construct a single trace for testbed replay in
the following manner.  For each WiFi measurement, we look at the most 
recent 3G measurement that precedes this WiFi measurement and repeat it
alongside the current WiFi measurement.  We do likewise for each 3G
measurement, thereby filling in the gaps to create a unified multi-network
trace.

For example, given the following independent 3G and WiFi measurements
(bandwidth in bytes/sec, rtt in ms):

            (3G) |   timestamp  |  bw_down | bw_up | rtt                                                                             
        -----------------------------------------------------                                                                        
                 |     1000.00  |    16800 |  7200 |  100
                 |     1003.00  |    32680 | 16000 |  150
        -----------------------------------------------------                                                                        
                                                                                                                                     
          (wifi) |   timestamp  |  bw_down | bw_up | rtt                                                                             
        -----------------------------------------------------                                                                        
                 |     1002.00  |   125000 | 32000 |  40
                 |     1005.00  |    96000 | 30000 |  70
                                                                                                                                     
The resulting trace will look like this:

              |         (wifi)        |           (3G)        |
   timestamp  | bw_down | bw_up | rtt | bw_down | bw_up | rtt |
   ------------------------------------------------------------
     1000.00  |    -    |   -   |  -  |   16800 |  7200 | 100 |
     1002.00  |  125000 | 32000 |  40 |   16800 |  7200 | 100 |
     1003.00  |  125000 | 32000 |  40 |   32680 | 16000 | 150 |
     1005.00  |   96000 | 30000 |  70 |   32680 | 16000 | 150 |
     

One effect of this post-measurement composition is that the downstream
bandwidth can be reported to be zero while the upstream bandwidth is
simultaneously non-zero.  This seems anomalous, but it is merely the result
of measuring downstream and upstream bandwidth sequentially rather than
concurrently.

To obtain traces with realistic network variability, we set up our testing
laptop in a vehicle and drove for extended time periods, continuously 
gathering measurements.  Using this setup, we obtained two traces in
different cities at different times of the day.

Trace #1 was gathered in Ann Arbor, MI in three distinct periods:
  11/23/2009
    1) 9:37-9:53pm EDT
    2) 11:09-11:25pm EDT
  11/24/2009
    3) 11:12am-12:58pm EDT
For ease of replay, we stitched these periods together by adjusting the
timestamps to close the gaps, forming one uninterrupted trace.

Trace #2 was gathered in Ypsilanti, MI in one continuous run, from
4:19pm-4:55pm EST, on 02/16/2010.


File Format
---
The format of the two provided trace files is similar to the above table.
Each file is a sequence of lines in this format:

  <timestamp> <wifi_down> <wifi_up> <wifi_RTT> <3g_down> <3g_up> <3g_RTT>

The first number is a UNIX-epoch timestamp of the testing completion time. 
The following three numbers describe WiFi downstream bandwidth, upstream
bandwidth, and round-trip time; the next three numbers describe the same
figures for the 3G network. All bandwidth figures are in bytes/second; all
round-trip times are in milliseconds.  For WiFi, if all three numbers are
zero, this signifies that WiFi was unavailable during that time.

Each line thereby represents a *slice* of the trace beginning at the 
timestamp, and the slices are sorted by timestamp.  For the purposes of
replay, in the last line of the trace, only the timestamp is used; this marks
the end of the last slice in the trace.


References
---
[1] Eriksson, J., Balakrishnan, H., and Madden, S. "Cabernet: vehicular
    content delivery using WiFi." In MobiCom 2008.

[2] Nicholson, A.J., Chawathe, Y., Chen, M.Y., Noble, B.D. and Wetherall, D.
    "Improved Access Point Selection." In MobiSys 2006.
