A hybrid direct-indirect estimator of network internal delays

The network tomography problem requires a remote source to estimate network-internal statistics on individual links within the network. Several techniques have been proposed to perform network delay tomography. The most accurate prior techniques directly measured delay by using ICMP Timestamp requests to the head and tail of the measured link. However, routing irregularities in the Internet significantly limited the applicability to individual links (the route to the head of the link had to be a proper prefix of the route to the tail of the link). Alternative techniques with broader coverage are all either less accurate, require new functionality in routers, or require the existence of a network-wide measurement infrastructure. This report presents a network tomography technique that achieves accuracy comparable with the best prior techniques, can measure one-way delays, depends only on existing infrastructure, and is much more widely applicable than previous techniques with comparable accuracy. Our approach is to use a number of probes to directly measure overlapping multi-link segments in the neighborhood of the target link. The delay distributions on these segments are used as input to an inference algorithm that derives an estimate of the queuing delay distribution on the target link. The coverage of this technique is assessed by inspecting many thousands of Internet paths. The accuracy is evaluated through simulation (where knowledge of the real delay distributions is obtainable, and where such real distributions can be compared to the estimated delay distributions).