Browse Theses

Accurately characterizing end-to-end Internet dynamics is exceptionally difficult, due to the network's immense heterogeneity. It can be impossible to gauge the generality of findings based on measurements of a handful of paths, yet logistically it has proven very difficult to obtain end-to-end measurements on larger scales.

At the heart of our work is a "measurement framework" we devised in which a number of sites around the Internet host a specialized measurement service. By coordinating "probes" between pairs of these sites we can measure end-to-end behavior along $O(N\sp2)$ paths for a framework consisting of N sites. Consequently, we obtain a superlinear scaling that allows us to measure a rich cross-section of Internet behavior without requiring huge numbers of observation points. 37 sites participated in our study, allowing us to measure more than 1,000 distinct Internet paths.

The first part of our work looks at the behavior of end-to-end routing: the series of routers over which a connection's packets travel. We find that pathologies increased significantly over the course of 1995, indicating that, by one metric, routing degraded over the year; that Internet paths are heavily dominated by a single route, but that routing lifetimes range from seconds to many days, with most lasting for days; and that, at the end of 1995, about half of all Internet paths included a major routing asymmetry.

The second part of our work studies end-to-end Internet packet dynamics.

To cope with such large-scaled measurements requires attention to calibration using self-consistency checks; robust statistics to avoid skewing by outliers; and automated "micro-analysis," such as that performed by tcpanaly, that we might see the forest as well as the trees.

With due diligence to remove packet filter errors and TCP effects, TCP-based measurement provides a viable means for assessing end-to-end packet dynamics.

We find wide ranges of behavior, so we must exercise great caution in regarding any aspect of packet dynamics as "typical."

Some common assumptions such as in-order packet delivery, FIFO bottleneck queueing, independent loss events, single congestion time scales, and path symmetries are all sometimes violated.

The combination of path asymmetries and reverse-path noise render sender-only measurements techniques markedly inferior to those that include receiver-cooperation. (Abstract shortened by UMI.)