Characterizing User Behavior and Network Performance ina Public Wireless LANAnand Balachandran Geoffrey M. Voelker Paramvir Bahl P. Venkat RanganU. C., San Diego U. C., San Diego Microsoft Research U. C., San Diego9500 Gilman Dr. 0114 9500 Gilman Dr. 0114 One Microsoft Way 9500 Gilman Dr. 0114La Jolla, CA 92093 La Jolla, CA 92093 Redmond, WA 98052 La Jolla, CA [email protected] [email protected] [email protected] [email protected] paper presents and analyzes user behavior and network perfor-mance in a public-area wireless network using a workload capturedat a well-attended ACM conference. The goals of our study are: (1)to extend our understanding of wireless user behavior and wirelessnetwork performance; (2) to characterize wireless users in terms ofa parameterized model for use with analytic and simulation stud-ies involving wireless LAN traffic; and (3) to apply our workloadanalysis results to issues in wireless network deployment, such ascapacity planning, and potential network optimizations, such as al-gorithms for load balancing across multiple access points (APs) ina wireless network.1. INTRODUCTIONAdvances in communication technology and the proliferation oflightweight, hand-held devices with built-in, high-speed radio ac-cess are making wireless access to the Internet the common caserather than an exception. Wireless LAN installations based onIEEE 802.11 [8] technology are emerging as an attractive solutionfor providing network connectivity in corporations and universi-ties, and in public places like conference venues, airports, shoppingmalls, etc. – places where individuals spend a considerable amountof their time outside of home and work. In addition to the con-venience of untethered networking, contemporary wireless LANsprovide relatively high data connectivity at 11 Mb/s and are easy todeploy in public settings.As part of a larger research project, we have been exploring issuesin implementing and deploying public-area wireless networks, andexploring optimizations for improving their performance [1]. Inorder to evaluate and validate the techniques that we are develop-ing, we consider it essential to use realistic workloads of user be-havior and wireless network performance to make design decisionsand tradeoffs. However, since public wireless LANs have only re-cently become widely deployed, such workload characterizationsare scarce. Initial studies of wireless networks have explored low-level error models and RF signal characteristics [5], installation andmaintenance issues of a campus wireless network [3], user mobil-ity in a low-bandwidth metropolitan area network [18], and userbehavior and traffic characteristics in a university department net-work [19] and, very recently, a college campus [11].In this paper, we extend previous studies by presenting and analyz-ing user behavior and network performance in a public-area wire-less network using a trace recorded over three days at the ACMSIGCOMM’01 conference held at U.C. San Diego in August 2001.The trace consists of two parts. The first part is a record of perfor-mance monitoring data sampled from wireless access points (APs)serving the conference, and the second consists of anonymizedpacket headers of all wireless traffic. Both parts of the trace spanthe three days of the conference, capturing the workload of 300,000flows from 195 users consuming 4.6 GB of bandwidth.The high-level goals of our study are three-fold. First, we want tosupplement the existing domain knowledge about wireless user be-havior and wireless network performance; research in Web infras-tructure, for example, has greatly benefited from the understandinggained from many workload studies from different settings overtime. By comparing and contrasting the workload in our settingwith previous ones, we can begin to identify and separate wirelessworkload characteristics that apply to the wireless domain in gen-eral from those that are specific to a particular setting or networkconfiguration. Second, we want to specifically characterize user be-havior and network performance in a public wireless LAN environ-ment. We characterize user behavior in terms of connection sessionlength, user distribution across APs, mobility, application mix, andbandwidth requirements; we characterize network performance interms of overall and individual AP load, and packet errors and re-transmissions. From these analyses, we characterize wireless usersin terms of a parameterized model for use with analytic and sim-ulation studies involving wireless LAN traffic. Third, we want toapply our workload analyses to better understand issues in wirelessnetwork deployment, such as capacity planning, and potential net-work optimizations, such as algorithms for load balancing acrossmultiple APs in a wireless network.For our conference workload trace, our overall analysis of user be-havior shows that:• In our setting, users are evenly distributed across all APs anduser arrivals are correlated in time and space. We can cor-relate user arrivals into the network according to a two-stateMarkov-ModulatedPoisson Process (MMPP). The mean inter-arrival time during the ON state is 38 seconds, and the meanduration of the OFF state is 6 minutes.• Most of the users have short session times: 60% of the usersessions last less than 10 minutes. Users with longer ses-sion times are idle for most of the session. The session timedistribution can be approximated by a General Pareto distri-bution with a shape parameter of 0.78 and a scale parameterof 30.76. The R2value is 0.9. Short session times implythat network administrators using DHCP for IP address leas-ing can configure DHCP to provide short-term leases, afterwhich IP addresses can be reclaimed or renewed.• Sessions can be broadly categorized based on their band-width consumption into light, medium, and heavy sessions:light sessions on average generate traffic at 15 Kbps, mediumsessions between 15 and 80 Kbps, and heavy sessions above80 Kbps. The highest instantaneous bandwidth demand is590 Kbps. The average and peak bandwidth requirements ofour users are lower than those in a campus network [19], re-flecting a difference in the type of tasks people do in the twosettings.• Web traffic accounts for 46% of the total bandwidth of all ap-plication traffic, and 57% of all flows. Web and SSH togetheraccount for 64% of the total bandwidth and 58% of flows.Our analysis of network performance shows that:• The load
View Full Document
Unlocking...