Related Work

There have been a number of studies on the access dynamics of web servers servicing clients over a wired network. These studies include analyses of web access traces from the perspective of proxies [7,20,21], browsers [6,9], and servers [4,16]. However, to our knowledge, all previous web workload studies have been conducted for browse services only and there are no published studies on notification services. Consequently, we believe, our analysis of notification services is the first study of its kind. Even for the browsing services, most studies analyze web servers serving clients over wired networks. There are very limited studies on web servers serving clients over wireless channels. The study closest to ours is the one done by Kunz et al. [12], which analyzes network traces generated by a mobile browser application. Specifically, their paper analyzes user behavior (bytes transferred and time spent on the wireless link) based on the notion of a session that was chosen to be 90 seconds; however, a different session period could potentially change their results. The main limitation of their work is the size of the data analyzed: although the traces were collected over a period of seven months, only 80K entries were logged. It is unclear whether the inferences drawn from this study can scale up to large commercial sites. In contrast, we analyzed traces with millions of entries generated over a period of 12 days at a large commercial site. Furthermore, their study also has the limitation that it uses client IP addresses for identifying users; since IP addresses can be reassigned to different users, it is difficult to perform an accurate user-based analysis. In our study, since every entry in the logs contains a unique identifier for every access/notification, we are able to carry out user-behavior analysis more accurately. In addition, our study is broader as we focus on user behavior, server load, content, and document popularity analysis. Tang and Baker analyzed a seven-week trace of a metropolitan-area packet radio wireless network, and a twelve-week trace of a building-wide local-area wireless network [18,19]. Both studies focus on how the networks were used, e.g., when the networks were most active, how active the network were, and how often users moved, etc. They did not consider the content or applications for which people used the wireless networks, which is the focus of our paper. Recently, Balachandran et al. [5] analyzed the user behavior and network performance of an IEEE 802.11 based wireless local area network (LAN) using a workload captured at a three day technical conference event. Their study focused on characterizing wireless LAN users for the purpose of coming up with a parameterized model to describe them. Additionally, they carried out workload analysis to address the network capacity planning problem. Their study is very different from ours in terms of analysis, methodology and objectives. While we focus primarily on wireless browse and notification services, they consider all network traffic for improving the network performance. Furthermore, the data-set they captured and analyzed is smaller and significantly different from the web server traces we analyze. In the sections that follow, whenever appropriate, we refer to related work done by other researchers and compare it with our findings.