Availability and Latency of World Wide Web Information Servers

During a 90-day period in 1994, we measured the availability and connection latency of HTTP (hypertext transfer protocol) information servers. These measurements 'were made from a site in the Eastern United States. The list of servers included 189 servers from Europe and 324 servers from North America. Our measurements indicate that on average, 5.0 percent of North American servers and 5.4 percent of European servers were unavailable from the measurement site on any given day. As seen from the measurement site, the day-today variation in availability was much greater for the European servers than for the North American servers. The measurements also show a wide variation in availability for individual information servers. For exampleo more than 80 percent of all North American servers were available at least 95 percent of the time, but 5 percent of the servers were available less than 80 percent of the time. The pattem of unavailability suggests a strong correlation between unavailability and geographic location. Median connection latency from the measurement site was in the 0.2-0.5 s range to other North American sites and in the 0.4-2.5 s range to European sites, depending upon the day of the week. Latencies were much more variable to Europe than to North America. The magnitude of the latencies suggest the addition of an MGET method to HTTP to help alleviate large TCP setup times associated with the retrieval of web pages with embedded images. The data show that 97 percent and 99 percent of all successful connections from the measurement site to Europe and North America respectively were made within the first 10 s. This suggests the establishment of client-side time-out intervals much shorter than those used for normal TCP connection establishment.