network. We examine a five month long log of user activity and traffic collected by a wireless network service provider operating hotspots in cafes, restaurants, ...
The 18th Annual IEEE International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC'07)
USER AND TRAFFIC CHARACTERISTICS OF A COMMERCIAL NATIONWIDE WI-FI HOTSPOT NETWORK Gautam Divgi and Edward Chlebus Network Modeling and Teletraffic Analysis Lab Department of Computer Science Illinois Institute of Technology Chicago, IL – U.S.A. Email: {divggau, chlebus}@iit.edu ABSTRACT This paper presents the first analysis of user and traffic characteristics in a commercial nationwide Wi-Fi hotspot network. We examine a five month long log of user activity and traffic collected by a wireless network service provider operating hotspots in cafes, restaurants, serviced apartments, hotels and airports all over Australia. We analyze traffic, user activity, user sessions and account usage. Several differences from previously studied Wi-Fi networks are observed and can be attributed to the user diversity and the commercial nature of the network. Key traffic and user statistics from the network are presented for reference. I.
INTRODUCTION
Wi-Fi technology originally designed to set up wireless LANs has become a foundation of ubiquitous wireless Internet access. Wi-Fi has made it available in airports, coffee shops, pubs, hotels, city centers, libraries, colleges etc. The wireless Internet market has recently grown dramatically and is still expected to attract more and more users due to deployment of new services and proliferation of portable wireless devices. Different types of wireless networks based on the IEEE 802.11 technology have been analyzed in literature until now. Tang and Baker analyzed the wireless network deployed in the Computer Science department at Stanford University [1]. Balachandran et al. studied the use of a wireless network at a conference [2]. Kotz and Essien studied a newly deployed wireless network at Dartmouth College [3] and Henderson et al. characterized the network again two years later [4]. Balazinska and Castro characterized a wireless corporate network spanning three buildings [5]. Blinn et al. analyzed the Verizon Wireless Hotspot Network (VWHN), which provides hotspot services on the island of Manhattan [6]. Traffic in a Wi-Fi network at two restaurants is briefly described in [7]. The various challenges in a wireless hotspot network have been examined in [8] and [9] proposes a business model for a wireless hotspot network. The above analyzed networks operate in academic [1, 3, 4], conference [2], corporate [5] or public environments [6, 7]. In these environments a user has unlimited freedom of access to the network, unlike a commercial Wi-Fi network where they operate under the constraints of account time and traffic volume limits. To the best of our knowledge no aspects of a commercial Wi-Fi hotspot network have been characterized in literature yet. Here, we present the first characterization of users and traffic in a commercial Wi-Fi network that serves diverse locations such as restaurants, serviced apartments, hotels and 1-4244-1144-0/07/$25.00 ©2007 IEEE.
airports all over Australia. We find significant differences in user behavior and traffic from previously studied wireless networks. In the next section we describe the measurement data, Section III defines various terms used in the paper, in Section IV we present our results and in Section V we conclude the paper and discuss possibilities for future work. II. MEASUREMENT DATA A. Description of the Study Environment The measurement data for this analysis were provided by Azure Wireless [10], one of the biggest wireless network operators in Australia. The network consists of access points in a number of public venues such as hotels, serviced apartments, restaurants, convention centers and airports. The user needs to have an IEEE 802.11 enabled device to access the network. This service is currently available at major cities in all the Australian states. Once customers are in the range of an Azure wireless hotspot they can connect to the network and purchase time to use it. To make this purchase users open their web browser on their 802.11 enabled device and try to access any valid website. The network then redirects them to an Azure Wireless login screen where they purchase time on the network by entering their credit card details and picking their account type from any of the two options below: • Hourly account: This type of account offers unlimited downloads and is billed per minute. The account is valid 1 year from the time of purchase, or until the time runs out. • Daily account: This type of account requires the user to enter the number of days of use and downloaded (inbound) data limit. The account is valid for the number of days chosen from the current date or until the data limit is reached. After completing the login process the user receives a login name and a password which can be used for any subsequent access to the network at any of its hotspot locations. As long as the account is active the user can log in any number of times. Wi-Fi sessions end when the users log out of the network. At the time of processing this data, the network operator imposed a 5 hour timeout on all sessions. B. Description of the Dataset Each entry in the log consists of 5 fields; the user name, the user login timestamp, the number of bytes inbound (downloaded) from the access point to the user’s device, the number of bytes outbound (uploaded) from the user’s device
The 18th Annual IEEE International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC'07)
to the access point and the duration for which the user was on the network. Users were not informed of the study so that their behavior would not be affected. To protect customer privacy all user specific information has been anonymized. The majority of the entries in the log correspond to accesses by customers. There are 688 testing and maintenance accesses generated by the network operator. These entries have been excluded from our analysis. A detailed model for wireless Internet session traffic variables is developed using these measurement data in [11]. III. DEFINITIONS To better understand user behavior and traffic in this network and for clarity in analyzing our results, we define the following terms and user classifications. Session: A session is created when a user logs into the network and ends when the user logs out or is timed out of the network. User: A user is the equivalent of a login name given to a purchaser of the network access time. A user can have multiple sessions until the account duration expires or the data limit is exceeded. Number of active days per user: The number of days in the log for which the user accesses the network. Inbound (downloaded) traffic: Traffic that flows from the access point into the users Wi-Fi enabled device. Outbound (uploaded) traffic: Traffic that flows from the users Wi-Fi enabled device into the access point.
The data received from Azure Wireless show a total of 14273 sessions generated over 153 days from 2nd October 2004 to 3rd March 2005. In that time all users generated 174.75 GB of traffic. The user and session activity per day over the entire duration of the log is shown in Fig. 1. From Fig. 1 we see a steadily increasing number of users not withstanding the sharp dip around the end of December and early January corresponding to the Christmas and New Year holiday period. The inbound and outbound traffic per day is shown in Fig. 2. From Fig. 2 we see that peaks in traffic do not mean a greater number of sessions for the same day in Fig. 1. There is a drop in traffic corresponding to the Christmas and New Year season; but when compared to session and user counts, traffic seems to be characterized more by sharp peaks occurring at irregular intervals. However, as a general trend traffic in the network shows a steady increase. B. User Behavior Since the user activity over the entire duration of the log is not uniform, we have analyzed daily and hourly user activity for February as it is the busiest month in the log. We can see from the hourly activity plotted in Fig. 3 that the time between 8 pm to midnight are the busiest hours. The daily activity in Fig. 4 shows Friday as the busiest day; however one must be careful in analyzing the results in Fig. 4 since they are calculated using only 4 sample points for each day of the week.
IV. RESULTS A. Overview of the Data
Figure 3: Average hourly session activity for February. The vertical bars represent the standard deviation. Busiest hours are between 8pm and midnight.
Figure 1: Session and user counts for each day across the duration of the log
Figure 4: Average session activity per day of the week for February (Only 4 points for each day of the week are available). The vertical bars represent the standard deviation
Figure 2: Inbound and outbound traffic volume for each day across the duration of the log
The hourly and daily user activity represent a significant departure from activity trends observed in academic [1, 3, 4], corporate [5] and public [6] wireless networks. A commercial Wi-Fi network shows busy usage during the night compared to a strong diurnal usage in other studies. It would seem that
The 18th Annual IEEE International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC'07)
most sessions generated in a commercial Wi-Fi network are for “after work” activity and not part of the users’ daily routine. For the daily activity we observe that unlike other studies where number of users dip sharply during the weekend, the number of users is mostly uniform for all the days of the week. From Fig. 5 we can see that the median number of distinct users per day is 40 which is just 1.93% of the total user population. Looking at the distribution for the number of active days per user in Fig. 6, we find that slightly less than half the users are present on only 1 day with 90% of the users present for less than 6 of the 153 days in the log. This is significantly lower than those observed in other studies. It would seem that users of a commercial Wi-Fi hotspot network access the network for casual use rather than everyday activity; also the low fraction of the total user population appearing per day shows a highly fluctuating user base.
by that fact that 90% of all users generate no more than 16 sessions over the lifetime of the account. The median session length is 24.62 minutes but roughly 10% of the users are connected to the network for more than 3 hours at a stretch. The distribution of the session length is given in Fig. 9. The session length has been successfully fitted to a general Pareto distribution in [2, 6] and is seen to follow a power law in [5]. We find that the sessions voluntarily terminated by the user (those not timed out by the network operator) fit the truncated Pareto distribution and can be attributed to the 5 hour timeout imposed by the operator. The fit of this distribution and the development of a model for key session traffic variables are examined in detail in [11].
Figure 8: Distribution of the number of sessions per user (The x-axis is truncated at 100)
Figure 5: Distribution of the number of sessions and users per day
Figure 9: Distribution of the session length (The x-axis is truncated at 19000 seconds)
Figure 6: Distribution of the number of active days per user (The x-axis is truncated at 40)
Figure 7: Distribution of the fraction of purchased account time used
We see from Fig. 7 that a user consumed a median of 76% of their purchased time and a little more than 20% of the users consumed all the time purchased by them. Fig. 8 shows that the median user generated 3 sessions. A few users of the network had very heavy usage, one user in particular generated 406 sessions. Such heavy use is rare and is borne
The network operator imposes a 5 hour timeout on all sessions but in 33 instances the imposition of this timeout seems to have failed. Those users whose sessions have been involuntarily terminated because of the timeout account for 5.6% of the total number of sessions. To measure how many users were active when timed out we computed the number of users who log back in less than 3 minutes after being involuntarily disconnected from the network. We find that of the 798 sessions that were timed out only 39 (