Load distribution of different users

In this section, we study the distribution of loads placed on the web server by different users. Our earlier analysis [1] examined the difference in load distribution between wireless users and offline users. Now we look at the load distribution at a more fine-grained level -- at a per-user level. Figure 16 and Figure 17 show the total number of accesses and total number of data requested by different clients, respectively (users with invalid identifiers were discarded). As the figures show, there is a significant variation in the load placed by different users on the web server: some users request several orders of magnitude more documents/data than other users. The accesses from only the wireless clients reveal similar property. Thus, service providers can consider designing different pricing plans that to cater to the widely varying needs of different users.

  

Figure 16: Total number of accesses made by different users.

  

Figure 17: Total number of data received by different users.

Figure 18 shows the inter-arrival time between requests coming from the same user. The requests generated from the offline users are much more bursty than those from the wireless users: 97% of the requests from the offline users have 1 second or less inter-arrival time, whereas only 9% of the requests from the wireless users have comparable inter-arrival time. We observe very bursty traffic for offline PDA users because their requests are generated by the downloader program rather than a human being; these users also generate significantly more requests than wireless users. If not handled appropriately, such bursts can delay wireless users unnecessarily. The web site designers can address this problem in a number of ways. For example, they can provide higher priority to wireless users or restrict the burst of offline user requests to a few front-door servers (servers that handle incoming HTTP requests). An orthogonal efficiency issue that needs to be addressed is the synchronization protocol for PDAs, i.e., instead of sending a large number of small requests, the synchronization protocol could batch all these requests into a single request and reduce the server load and roundtrip latency.

  

Figure 18: CDF of inter-arrival time between consecutive requests from the same user.