We Are Extremely Grateful to Dynamic Server Selection Using Bandwidth Probing in Wide-area the Bsd Packet Filter: a New Architecture for User-level 8. Previous Work 8.1 Network Characterization 8.2 Network Probing 7.2 Methodology

[22] Official 1996 olympic web site-home page. In this paper, we have presented a statistical analysis of through-puts of Web transfers. This analysis was performed using a large packet-level trace of traffic at the Web site for the Atlanta Summer Olympic Games. Observed mean throughputs for these transfers vary widely as a function of end-host location and time of day, confirming that a large amount of heterogeneity exists in the Inter-net. Our analysis showed that despite these wide variations the transfer throughput exhibited significant temporal and spatial stability. In particular, we found that: 1. Throughputs to several individual internet hosts can be mod-eled using a log-normal distribution regardless of location. More precisely, we found that for over 50% of the hosts in our study, the log-normal hypothesis could not be rejected at any level of significance, and that it characterized observed throughputs better than the other analytic distributions we considered. 2. Nearby Internet hosts often have almost identical probability distributions of observed throughput. The size of the clusters for which performance is identical varies as a function location on the Internet. However, our analyses show that cluster sizes in the range of 2-4 hops work for many regions. 3. Our analysis of a small number of hosts shows that end-to-end throughputs to hosts often vary by less than a factor of two over time scales on the order of many tens of minutes. We also find that throughputs are piecewise stationary over time scales of this magnitude. Overall, these traffic characteristics indicate that there is great promise in techniques that cache and share such network characteristics amongst nearby hosts as well as within a single host across connections. Schwithenberg, and others who did an admirable job of running the Atlanta Olympics Web server at the site in Southbury, CT, for their support and help. They provided patient help and advice while under enormous pressures in getting the Web site up and running. We also thank Arvind Krishna and Hamid Ahmadi for their support and encouragement, as well as comments on earlier drafts of this paper. We would also like to thank the anonymous SIGMETRICS reviewers for their detailed comments, suggestions, and criticisms, that led to significant improvements in the quality of this paper. Obviously, one cannot generalize these results to make general statements about the stability of Internet performance, but it is a start. In particular, the fact that performance …