A critical first step to improving existing and designing future wide-area networks is an understanding of the load placed on these networks. Efforts to model traffic are often confounded by atypical traffic - traffic particular to the observation site not ubiquitously applicable. The causes and characteristics of atypical traffic are explored in this thesis. Atypical traffic is found to interfere with parsimonious analytic traffic models. A detection and modeling technique is presented and studied for atypical traffic characterized by strongly clustered inliers. This technique is found to be effective using both real-world observations and simulated data.

Another form of atypical traffic is shown to result in multimodal distributions of connection statistics. Putative methods for bimodal estimation are reviewed and a novel technique, the midpoint-distance profile, is presented. The performance of these estimation techniques is studied via simulation and the methods are examined in the context of atypical network traffic. The advantages and disadvantages of each method are reported.