Improving Web experience by client characterization driven server adaptation

In this paper we seek to categorize the set of clients communicating with a server on the Web based on information about these clients that can be determined by the server. We use this information to help drive tailored actions on the part of Web servers. Users with poor connectivity may choose not to spend much time at a Web site if it takes a long time to receive a page, even if the Web server at the site is not the bottleneck. Retaining such clients may be of interest to a Web site. Better connected clients may be able to receive enhanced representations of Web pages such as with higher quality images. Our work explores a variety of considerations that could be used by a Web server in characterizing a client. Once a client is characterized as poor, a variety of server actions can be taken by the server such as delivering altered content, altering how content is delivered, guiding server policy decisions, aiding in caching decisions or deciding when to redirect the client to a mirror site. We also use network-aware client clustering techniques to provide a coarser level of client categorization. Such a cluster categorization can then be used to categorize subsequent clients from that cluster for which a client-speciic categorization is not available. Results for client characterization and applicable server actions are derived from real, recent, and diverse set of Web server logs. Our experiments demonstrate that a relatively simple characterization policy can classify \poor" clients such that these clients subsequently make the majority of poor performing requests to a Web server. This policy is also stable in terms of clients staying in the same class for a large portion of the analysis period. Client clustering can signiicantly help in initially classifying clients for which no previous information about the client is known. We also show Improving Web experience by client characterization driven server adaptation 2 that diierent server actions can be applied to a signiicant number of request sequences with poor performance. Improving Web experience by client characterization driven server adaptation