Measuring wireless network usage with the

Experience Sampling Method

Tristan Henderson Denise Anthony David Kotz
Department of Computer Science Department of Sociology Department of Computer Science
Dartmouth College Dartmouth College Dartmouth College
Hanover, NH 03755, USA Hanover, NH 03755, USA Hanover, NH 03755, USA
[email protected] [email protected] [email protected]

Abstract— Measuring wireless local area networks has proven We introduce the Experience Sampling Method in the next
useful for characterizing, modeling and provisioning these net- section, and describe previous ESM studies and wireless
works. These measurements are typically taken passively from a network measurement studies. Section II describes our exper-
vantage point on the network itself. Client devices, or users, are
never actively queried. These measurements can indicate what imental setup and methodology, and Section III presents some
is happening on the network, but it can be difficult to infer brief results. In Section IV we list some of the lessons that
why a particular behavior is occurring. In this paper we use the we learnt in the course of conducting the study, and Section V
Experience Sampling Method (ESM) to study wireless network concludes.
users. We monitored 29 users remotely for one week, and signaled
them to fill out a questionnaire whenever interesting wireless I. BACKGROUND AND RELATED WORK
behavior was observed. We find ESM to be a useful method
for collecting data about wireless network usage that cannot Diary methods are a group of research tools and meth-
be provided by network monitoring, and we present a list of ods used in psychology for “documenting the particulars
recommendations for network researchers who wish to conduct of life.” [6]. The distinguishing feature of a diary method
an ESM study. is that participants self-report their own ongoing experi-
Researchers have conducted many wireless local area net- ences. This enables the recording of events and experiences
work (WLAN) measurement studies, collecting data to help in a more natural context than a formal interview. Diary
characterize, model and provision these networks. Typically methods can be divided into three categories [22]: interval-
these studies take network-side measurements, counting the contingent, where participants report at regular intervals;
number of clients, the amount of traffic transmitted, client signal-contingent, where participants report when they receive
session durations and so forth. Such studies are useful for a signal; and event-contingent, where participants report when-
determining what is occurring on a network. They may not, ever a defined event occurs.
however, explain why a particular behavior is observed. For in- Within the broader category of diary methods, the Experi-
stance, in one of our recent studies [13], we observe increased ence Sampling Method (ESM) [18] has emerged as a popular
peer-to-peer filesharing on our campus WLAN, despite the method for evaluating user experiences and situations. In an
presence of a higher-bandwidth wired network. Why are users ESM study, also referred to in the literature as Ecological
choosing to use the slower wireless network for downloading Momentary Assessment (EMA), participants fill out a ques-
such large files? Is it because the convenience of the wireless tionnaire several times a day. The questionnaire asks about
network outweighs its limited capacity? Or is it because the the participant’s current activities, conditions and feelings.
users are unaware of the wired network? Such questions cannot A typical ESM study is signal-contingent, involving seven
be answered by network-side measurement alone. signals a day over seven days. ESM has been used to evaluate
a wide variety of experiences, such as job satisfaction amongst
To understand why particular wireless network behavior
teachers [5], happiness amongst American youths [9], and
occurs, it is useful to ask network users about their usage pat-
the quality of experience of new parents before and after
terns. A one-off survey or questionnaire, however, may not be
childbirth [10].
appropriate. If a user is asked about a wireless behavior some
Most relevant to our work are ESM studies that look
time after the event has occurred, they may have forgotten the
specifically at communications or technology. Kubey et al. [17]
conditions at the time of the event, or the particular reason
present a survey of communications-related ESM studies, such
for their actions. Instead, we need to way to non-intrusively
as television viewing, and communication patterns within
record all aspects of a user’s wireless network experience.
families and marriages. Consolvo and Walker [8] use ESM
In this paper, we describe an Experience Sampling Method to evaluate user experiences with the Intel Personal Server
(ESM) study of a campus-wide WLAN. ESM is a psychology ubiquitous computing device. Gaggioli et al. use ESM to
method for studying user experience. We compare the data analyze quality of experience and presence in virtual envi-
obtained through ESM to data from network monitoring. ronments [12].
 Technology can also be used to improve ESM methodol- were extensive, containing 25 questions, some of which had
ogy. Instead of using a notebook or paper questionnaires to several parts and open-ended questions, and thus would be
record responses, Palen and Salzman ask their participants impractical to fill out using a PDA. Second, we considered
to record voicemail messages upon receiving an alert [19]. using a PDA equipped with an 802.11 network adapter, as this
Another common electronic ESM tool is to use a PDA for would have allowed us to use the same device for signalling
questionnaires [4]. participants, for participants to fill out questionnaires, and
Several researchers have measured and characterized wire- for the questionnaire responses to be transmitted to a central
less LANs, mainly in academic settings. Academic WLANs data-collection server. We rejected this solution, however, as
that have been measured include those at Stanford [21], participants could not be signalled when they were out of
Georgia Tech [14], Dartmouth [13], [16], UNC [7] and the range of an 802.11 network (for example, if they were off
University of Saskatchewan [20]. These studies have all taken campus). Moreover, participants could not be signalled if the
place on the wired side of the access points, using packet 802.11 network was having trouble, which was one of the
sniffers, SNMP polling, syslog records, or a combination of aspects that we were interested in monitoring. Rather than
these techniques. The non-academic WLANs that have been require participants to carry two electronic devices (a PDA
studied include a conference WLAN [2], and the WLAN at and a pager), which would only increase the probability of a
a corporate research campus [3], both of which were also device malfunction during the study, we chose to use a pager
measured using SNMP or packet sniffing. and a notebook.
More recently, WLANs have been measured from the wire- Participants provided “conflict times” at which they did not
less side of the access point; that is, measuring the wireless want to be paged, for example, when they were asleep or in an
medium itself. Wireless-side measurement is complex due examination. Using this information, we drew up a schedule of
to the unreliable nature of the wireless medium, hardware signal-contingent alerts. At each signal-contingent alert time,
and software incompatibilities, and so forth. As such, most we sent participants a page containing a numeric code. This
wireless-side studies are of a smaller scale than wired-side code indicated the page in the notebook to be filled out (in case
studies. Yeo et al. measure a single access point at the a participant missed an alert). Each participant received seven
University of Maryland [23], while various other studies have signal-contingent alerts per day, and each alert was scheduled
taken place in controlled wireless conditions [1], [11], [15]. at least 45 minutes after the preceding alert, to prevent the
alerts from being too intrusive.
II. M ETHODOLOGY
B. Wireless monitoring setup
In this section we describe our experimental setup. Our
study differs slightly from most ESM studies in that it is both We monitored our campus WLAN using our existing mea-
signal-contingent and event-contingent. Participants are sent surement infrastructure. This uses four techniques to trace
signals to record details about their general wireless network WLAN usage: syslog and DHCP records, SNMP polls and
experience, but in addition, we use our network monitoring network sniffers (see [13] for further details). We record
infrastructure to detect interesting events and send event- syslog messages from every AP on campus every time a
contingent signals. user associates, disassociates, authenticates, deauthenticates or
roams. Every five minutes, we poll each AP using SNMP to
A. ESM setup collect detailed client statistics including bytes, packets and
We recruited participants using a website and college bul- errors transmitted; signal strength and quality. We have in-
letin boards. The website used a registration form to preselect stalled 18 Linux boxes in various switchrooms around campus,
participants (to make sure, for instance, that all participants connected to the wired side of the APs, and use tcpdump to
owned a wireless device). We interviewed all of the respon- capture headers of packets sent to and from the APs.
dents and selected 30 participants: 15 male and 15 female. In addition to the measurement mechanisms described
All of the participants were undergraduate students. To attract in [13], we obtained logs from our central campus e-mail
students and encourage participation, we offered them $100 server. This server writes a timestamped record to a log
each to participate in the study. One participant dropped out whenever a user logs on to the e-mail server (logout times are
during the study; this participant’s data is excluded from the not recorded). Our system administrators filtered these logs by
results. username, to provide us with records for our 30 participants.
We issued each participant with a pager and a questionnaire To detect interesting wireless events, we used the syslog and
notebook. The pager was a Motorola Bravo numeric pager SNMP data since these are collected in real-time.1 We detected
operating in the 406-512MHz band. To programmatically send the following events (the data source is listed in brackets):
pages to the participants, we used a SMTP-to-page gateway • [syslog] Excessive associations and disassociations in a
operated by the pager service provider. The notebook con- short time period (“ping-ponging” between APs). We define
tained questionnaires with questions about communications ping-ponging as a median interarrival between a client’s last
usage (wired, wireless and other communications devices), five associations of less than 10 seconds.
their current activities, location and so forth. We chose to 1 While the sniffers collect data in real-time, logs are stored on the sniffers
use a notebook for several reasons. First, our questionnaires and uploaded to our analysis server in off-peak periods.
• [SNMP] A client associates with an AP that is heavily- Fig. 1. Number of locations at which a participant is signalled, compared
loaded with clients. We define such a “busy” AP as one to the number of locations where a participant is signalled and connected to
the wireless network. Distribution across locations per day.
with more than 16 associated clients. 1
• [SNMP] A client associates with an AP that is not reachable
0.9
via SNMP. An AP that cannot be reached by SNMP, but
0.8
is still sending syslog messages, may be having network or
0.7
operational problems. We define such an “unreachable” AP
0.6
as one that has not been polled successfully for one hour.
0.5
• [SNMP] A client associates with an AP that is heavily-
0.4
loaded with traffic. We define such a “high traffic” AP as
0.3
one where the mean inbound or outbound bytes per second
0.2
over the last two SNMP polls is greater than 250Kb/s.
0.1 All locations
• [SNMP] A client associates with an AP that is experiencing Locations when on wireless
0
high levels of errors or retransmits. We define such an 0 1 2 3 4 5 6
“error-prone” AP as one with an error/retransmit level of Daily number of locations
≥ 10%.
• [syslog] A client associates with an AP after a long period
of inactivity, defined as two hours. We ignored the first
association of the day, as this might be the participant correctly-determined locations is surprisingly low. It means
waking up. that our raw syslog data may require some processing or ad-
Whenever one of the above events occurred, an event- ditional context-specific information before conclusions about
contigent alert was triggered, which sent a page to the relevant user locations or user mobility can be made.
participant. These pages contained a different code, indicating
that the participant should fill out an event-contingent survey. The ESM questionnaires provide a useful mechanism for
This survey was shorter, containing questions that were solely understanding errors in location-determination. Where there
about the participant’s wireless experience. is a discrepancy in locations, syslog reported a building that
Each participant was sent up to three event-contigent alerts is adjacent to the self-reported location 5% of the time. Errors
a day. Event alerts were only sent if the participant had not in the self-reporting, such as a blank entry in the location field
received a page in the last 45 minutes. of the questionnaire, account for 13.4% of the discrepancies.
The study took place over one week, starting on a Sunday. The largest cause of discrepancies is where we believe the
participant leaves their laptop at home, but still powered up
III. R ESULTS and connected to the wireless network. We distinguish this
In this section we present some results from our study. behavior by looking for a long wireless session, where a client
The study generated an enormous amount of data; with 29 is observed associated with the same AP, but the corresponding
users filling out up to 10 questionnaires a day over seven participant’s self-reported locations change more than once.
days, there were over 270,000 individual data points from the These situations account for 23.2% of the discrepancies.
questionnaires alone, in addition to the data from pre-study and One use for tracking the location of network users is
post-study interviews, and the network monitoring. As such, in to estimate or model their mobility. Using AP associations,
this paper we only provide selected statistics, and concentrate however, only allows us to track user mobility while they are
on those results that could not be obtained through network- associated with the network. While this may be appropriate
monitoring. for laptop users, who are nomadic in their network usage,
A. User location newer wireless devices such as VoIP phones tend to be always-
on, and so have different mobility characteristics. To better
In previous measurement studies we have used syslog
anticipate the nature of these devices’ mobility, we may want
records to determine the approximate location of a user at
to track users’ overall physical mobility, i.e., track all of the
a given point in time. We use a client’s association records,
locations that they visit, rather than just those where they are
and assume that the client is located near to this AP. This
using the WLAN.
location information may not always be accurate, however, due
to different client driver behavior, or ping-ponging between Figure 1 shows the total number of different locations at
APs, which can make it difficult to determine an accurate which a participant self-reported each day, compared to the
location. Asking users to self-report their location may solve number of locations at which they self-reported themselves
these problems. and were also connected to the wireless network. The median
Comparing the locations determined via syslog with the total number of locations is 2, whereas the median number of
participants’ self-reported locations at the time of a page shows wireless locations is 0. By only observing users on our campus
that syslog accurately determined the building in which a while they are connected to the WLAN, we are clearly missing
participant is located 55% of the time. This proportion of out on a large part of their mobility.
B. Device usage Fig. 2. Total participant traffic (GB), by TCP or UDP protocol.
By monitoring an 802.11 wireless network, we can observe Inbound
Outbound
the behavior of the clients on that network. By analyzing TCP database
flows, it may also be possible to determine the device type of bulk
each client, which we have done in a previous study [13]. We interactive

are limited, however, to wireless devices that use the 802.11 unknown

PHY and MAC layers. A given user may be carrying a variety services

of other wireless devices, such as cellphones or Bluetooth streaming

devices. It is useful to understand how these multiple devices mail

are used, to influence the design of future hybrid network filesystem

devices, such as phones that can operate on both 802.11 and www

UMTS networks. p2p

We asked participants about all of their communications, 0 1 2 3 4 5 6 7 8 9

both electronic and non-electronic, and the devices that they Total traffic (GB)
used, at the time of an alert. Laptops were the most commonly-
used devices, in use at 28.8% of alerts. The television was
the next most popular device, in use at 10.3% of alerts.
Desktops and public terminals were used at 1.9% and 2.9% of participants to self-report their application usage means that
alerts respectively. Other devices in use included MP3 players we can examine their usage of both wired and wireless
(1.3%), cellphones (1.0%) and analog phones (0.9%). networks.
Participants were often using more than one device. The
television, in particular, was used in conjunction with other Examining DHCP records for participants’ wired and wire-
devices; 48.3% of the events that involved television usage also less MAC addresses shows that of our 29 participants, 13
involved the use of another device. 23.7% of the events that connected to the wired network at some point during the week.
involved laptop usage also involved another device; of these All participants connected to the wireless network.
multi-device events, 34.1% involved television and laptops, Figure 2 shows the overall application types that we saw
and 12.7% involved MP3 player and laptops. on our tcpdump wireless sniffers. We use well-known port
Dartmouth has gradually migrated from an analog PSTN numbers to identify applications, and in addition we parse
telephony system to a VoIP system over the past year. Wireless data payloads to look for some of the more popular peer-to-
VoIP devices (Cisco 7920 and Vocera badges) are deployed peer filesharing programs (DirectConnect, BitTorrent, Kazaa
on campus, and students are provided with Cisco Softphone and so forth; see [13] for further methodology details). It
VoIP software for use on their laptops. We have found very is clear from Figure 2 that such peer-to-peer filesharing
little wireless VoIP usage so far, however [13]. This was accounted for the majority of the traffic generated by our
confirmed in our ESM study; only one participant used VoIP. study participants (73.0%). At first glance this is not reflected
We also found very little cellphone usage, and participants in our questionnaires, where participants report that they are
reported using a cellphone in the hour prior to an alert at using a peer-to-peer application at only 2.6% of the alerts, and
only 2.1% of the alerts. There is limited cellphone coverage only six participants report using a peer-to-peer application at
on our campus, due to its somewhat isolated location. As all. Further traffic analysis indicated that most of the peer-to-
such, students tend to prefer using e-mail communication peer traffic (96.4%) was generated by four participants, one of
over telephony; participants reported sending e-mail in the 30 whom never reported this usage. The timing of the alerts and
minutes before an alert at 15.5% of alerts, with an average the location of our packet sniffers, however, meant that we did
of 4.7 e-mail messages sent at these times. Our low observed not capture all peer-to-peer usage, with two users reporting
VoIP usage might therefore be due to low telephone usage in peer-to-peer usage off-campus. ESM can thus complement
general. packet sniffers; ESM can indicate application usage in areas
C. Network usage that are difficult to monitor, and packet sniffing can be used
to verify the accuracy of ESM reports.
Our campus measurement studies have concentrated on the
wireless portion of the campus LAN. To understand trends in The ESM questionnaires are also useful for determining the
wireless network usage, it may be useful to examine usage on purpose of network traffic. For instance, if participants were
the wired network as well, to determine those behaviors that using a web browser at the time of an alert, we asked them
are wireless-specific. Measuring our wired network is difficult, whether they were surfing for work-related purposes, pleasure,
due to the lack of convenient central points to monitor traffic, or both. When participants were not at home, they surfed for
and the large amounts of data involved. Moreover, both our work 56.9% of the time and pleasure 54.2% of the time. When
wired and wireless networks do not require authentication, and at home, however, the level of pleasure-related surfing rose
so it is impossible to determine which wired and wireless significantly to 74.1%, while participants surfed for work at
MAC addresses represent the same client machine. Asking 30.2% of alerts.
 IV. L ESSONS FROM THE STUDY recently disconnected from the wireless network, but over
five minutes ago. We may have been able to obtain more
This was the first ESM study to be conducted by the authors, data if we had asked participants about their wireless activity
and one of the first ESM studies to involve the interactive use over a longer timescale, such as the last 30 minutes.
of a wireless network. We have learnt several lessons from this • Be flexible about user preferences.
initial study, and we list some of these for other researchers In our initial testing, we assumed that participants would
who may wish to carry out an ESM study. have somewhat static conflict times, so that it would be
• Choose the time of the study carefully. possible to predetermine all of the times for the signal-
ESM studies typically last for one week, as it is difficult to contigent alerts. It turned out that students have more
ensure participation for longer periods. When holding a one- unpredictable schedules than we had assumed, and many
week long study, it is important to choose an appropriate participants contacted us during the week of the study to
week. We chose a week during the summer term, as this was modify their conflict times. Some students, in fact, seemed
convenient for the researchers and faculty members who to think that it was reasonable to modify their conflict
were conducting the study. Unfortunately there are fewer times with only five minutes warning! We had to quickly
students on our campus during this term, which meant that modify the program that sent out the signal-contigent alerts
there was a smaller pool of students to recruit from, and to adapt to changing schedules. For future studies, however,
that there were fewer users on the network. In addition, we recommend instituting a rule that participants must
some of our participants had examinations during the week provide sufficient notice (e.g., 12 hours or 24 hours) before
of the study, and they commented that they did not use rescheduling their conflict times.
their wireless devices as much as usual because they were • Clearly explain the technical parameters of the study.
studying for their examinations. The study was described to participants as a study about
• Allow sufficient time between participant selection and the their communication patterns and wireless network usage.
start of the study. This non-technical description was not specific enough.
The timing of participant selection for a sociology study During our study, one participant contacted us to say that
is a complicated procedure. If the time interval between they had left campus and gone home (to another state).
participant selection and the start of a study is too long, They thought that they could still participate in the study,
then the participants may forget to participate, or forget their because they were using their own WLAN at home. We had
instructions, by the time the study begins. In our study, the to explain that we could only monitor network behavior on
users were selected on Friday, and the study began on a our campus WLAN.
Sunday. We found that this was insufficient time to prepare • Automatically prescreen participants.
the study, due to the amount of data that needed to be In our study the only automated prescreening was through
collected from all of the participants before the study could the use of a webpage, where prospective participants had
begin. For instance, several participants provided us with to fill out contact details and state whether they owned
incorrect MAC addresses, that required verification. Several a wireless laptop. Additional prescreening took place in
participants gave us too many or overlapping conflict times, face-to-face interviews. It might, however, be preferable to
so that it was impossible to schedule seven signal-contingent increase the level of automated prescreening. For instance,
alerts. the webpage could ask for a user’s MAC address, and the
• Consider all event-contingent alerts. network usage of this MAC address could be monitored
Our study resulted in very few event alerts being sent to prior to interviewing the prospective participant. This could
participants. This was due to a number of factors. Holding be used to ensure participants with specific home locations
the study during the summer term meant that there were or usage patterns.
fewer users on the network, and this may have led to less • Use a two-way signalling channel.
problems of network congestion or loss. Although we asked The biggest practical problem with our study was the radio
our participants to minimize their list of conflict times, it pagers used to signal the participants. We chose to use
turns out that there was insufficient free time in each day pagers as they provided a simple interface for participants,
to send participants three event-contingent alerts each day, and as they would still function even when out of range
in addition to the seven scheduled signal-contigent alerts. of our campus WLAN. Unfortunately our pagers were
This was because of our requirement that each alert be sent one-way devices; they could receive pages, but the device
at least 45 minutes apart. With hindsight, we should have itself had no means of transmitting any data. Thus there
scheduled the signal-contingent alerts at least 45 minutes was no way to receive feedback to determine whether a
apart, but sent the event-contingent alerts whenever we message had been successfully sent. In practice we found
observed one of the event-triggering network behaviors. that some participants did not receive all of their pages,
• Provide a sufficiently large time window for responses. or that their pages were queued by the SMTP gateway
Our questionnaires asked participants about their activities and received during a conflict time. For future studies we
in the five minutes previous to receiving a page. Some recommend the use of a two-way pager, so that participants
users often mentioned in their responses that they had can acknowledge receipt of a signal.
• Maintain an anonymous support channel. [3] M. Balazinska and P. Castro. Characterizing Mobility and Network
In any study that involves the collection of data from Usage in a Corporate Wireless Local-Area Network. In Proceedings of
the First International Conference on Mobile Systems, Applications, and
human subjects, it is vital to maintain participants’ privacy. Services (MobiSys), pages 303–316, San Francisco, CA, May 2003.
In an academic environment, this may be a requirement [4] L. F. Barrett and D. J. Barrett. An introduction to computerized
of the Institutional Review Board (IRB) or equivalent. To experience sampling in psychology. Social Science Computer Review,
19(2):175–185, Summer 2001.
maintain privacy, we assigned anonymous identifiers to each [5] A. Bishay. Teacher motivation and job satisfaction: A study employing
participant and used these instead of participants’ names in the experience sampling method. Journal of Undergraduate Science,
the signalling program. The identifier to name mapping was 3:147–154, Fall 1996.
[6] N. Bolger, A. Davis, and E. Rafaeli. Diary methods: Capturing life as
kept in a separate location to the researcher that operated it lived. Annual Review of Psychology, 54:579–616, Feb. 2003.
the signalling program, so that a security breach would not [7] F. Chinchilla, M. Lindsey, and M. Papadopouli. Analysis of wireless
result in the revelation of any private information. There information locality and association patterns in a campus. In Proceedings
of the 23rd Annual Joint Conference of the IEEE Computer and
may, however, be situations where participants need to Communications Societies (INFOCOM), pages 906–917, Hong Kong,
contact researchers, and by doing so they may reveal their China, Mar. 2004.
identities. Several participants had support issues, such as [8] S. Consolvo and M. Walker. Using the experience sampling method to
evaluate ubicomp applications. IEEE Pervasive Computing, 2(2):24–31,
pagers failing to work, or needed to contact us to reschedule Apr.-June 2003.
their conflict times. Providing an online bulletin board or [9] M. Csikszentmihalyi and J. Hunter. Happiness in everyday life: The uses
mailing list, where participants can communicate using of experience sampling. Journal of Happiness Studies, 4(1):185–199,
2003.
anonymous identifiers, is a useful means for maintaining [10] A. D. Fave and F. Massimini. Parenthood and the quality of experience
participant anonymity. in daily life: A longitudinal study. Social Indicators Research, 67(1-
2):75–106, June 2004.
V. C ONCLUSIONS AND F UTURE W ORK [11] G. Gaertner and V. Cahill. Understanding link quality in 802.11 mobile
ad hoc networks. IEEE Internet Computing, 8(1):55–60, Jan/Feb 2004.
In this paper we present some results from an Experience [12] A. Gaggioli, M. Bassi, and A. D. Fave. Quality of experience in virtual
Sampling Method study of a campus-wide wireless network. environments. In G. Riva, F. Davide, and W. A. Ijsselsteijn, editors,
Being There: Concepts, effects and measurement of user presence in
Our study is novel in that it is the first to integrate network synthetic environments, pages 121–135. Ios Press, Amsterdam, The
monitoring into an event-contingent study, so that events can Netherlands, 2003.
be triggered by observed network behavior. [13] T. Henderson, D. Kotz, and I. Abyzov. The changing usage of a
mature campus-wide wireless network. In Proceedings of the Tenth
We find that an ESM study is a useful method for augment- Annual International Conference on Mobile Computing and Networking
ing the data collected from a traditional network monitoring (MobiCom), pages 187–201, Philadelphia, PA, Sept. 2004.
study. Conducting an ESM study, however, is a complex [14] R. Hutchins and E. W. Zegura. Measurements from a campus wireless
network. In Proceedings of the IEEE International Conference on
process, given the interactions with people and the large Communications (ICC), volume 5, pages 3161–3167, New York, NY,
number of variables that can cause things to go wrong. By Apr. 2002.
presenting the lessons that we have learnt from our study, we [15] A. Kochut, A. Vasan, A. U. Shankar, and A. Agrawala. Sniffing out
the correct physical layer capture model in 802.11b. In Proceedings of
hope that other researchers can benefit from our experiences. the 12th IEEE International Conference on Network Protocols (ICNP),
In future work we intend to improve our analysis methods, pages 252–261, Berlin, Germany, Oct. 2004.
in particular allowing us to better correlate data from ESM [16] D. Kotz and K. Essien. Analysis of a campus-wide wireless network.
Wireless Networks, 11:115–133, 2005. Previously appeared in MobiCom
questionnaires and network monitoring. We also intend to 2002 and as Dartmouth TR2002-432.
carry out further ESM studies, using the lessons that we have [17] R. Kubey, R. Larson, and M. Csikszentmihalyi. Experience sampling
learnt from this study. method applications to communication research questions. Journal of
Communication, 46(2):99–120, June 1996.
VI. ACKNOWLEDGEMENTS [18] R. Larson and M. Csikszentmihalyi. The experience sampling method.
New Directions for Methodology of Social and Behavioral Science,
This work was a project of the Center for Mobile Computing 15:41–56, 1983.
at Dartmouth College, and was supported by McKinsey & [19] L. Palen and M. Salzman. Voice-mail diary studies for naturalistic data
capture under mobile conditions. In Proceedings of the ACM Conference
Company, the Cisco Systems University Research Program on Computer Supported Collaborative Work, pages 87–95, New Orleans,
and NSF Infrastructure Award number EIA-9802068. The LA, Nov. 2002.
authors thank Linda Lomelino ’06 for her help with screen- [20] D. Schwab and R. Bunt. Characterising the use of a campus wireless
network. In Proceedings of the 23rd Annual Joint Conference of the
ing participants and processing the questionnaire data, and IEEE Computer and Communications Societies (INFOCOM), pages
WaveComm of West Lebanon, NH, for providing the pagers. 862–870, Hong Kong, China, Mar. 2004.
We also thank Minkyong Kim and the anonymous reviewers [21] D. Tang and M. Baker. Analysis of a local-area wireless network. In
Proceedings of the Sixth Annual International Conference on Mobile
for their helpful comments on earlier drafts of the paper. Computing and Networking (MobiCom), pages 1–10, Boston, MA, Aug.
2000.
R EFERENCES [22] L. Wheeler and H. T. Reis. Self-recording of everyday life events:
[1] A. Aguiar and J. Klaue. Bi-directional WLAN channel measurements in Origins, types and uses. Journal of Personality, 59(3):339–354, Sept.
different mobility scenarios. In Proceedings of the Vehicular Technology 1991.
Conference (VTC Spring), Milan, Italy, May 2004. [23] J. Yeo, M. Youssef, and A. Agrawala. A framework for wireless LAN
[2] A. Balachandran, G. M. Voelker, P. Bahl, and P. V. Rangan. Character- monitoring and its applications. In Proceedings of the Third ACM
izing user behavior and network performance in a public wireless LAN. Workshop on Wireless Security (WiSe’04), pages 70–79, Philadelphia,
In Proceedings of the 2002 ACM SIGMETRICS Conference, pages 195– PA, Oct. 2004.
205, Marina Del Rey, CA, June 2002.