Direct measurement vs. indirect inference for determining network-internal delays

We consider direct techniques for measuring network-internal delays that rely on network routers as measurement nodes, and indirect techniques that rely on end-to-end measurements and infer delay distributions on the common internal paths. The direct technique does not require any additional infrastructure support, and can thus be directly applied to the analysis of network performance and the study of the Internet. The use of routers for measurement also results in a simple mathematical formulation of the delay estimation problem, thereby providing robustness and accuracy. However, the irregularity of Internet routing somewhat limits the applicability of the direct technique. In contrast, indirect techniques are not sensitive to routes to internal nodes, but accumulate increased errors over longer paths. We use simulation to compare the direct and indirect methods and present results from network measurements exploring the applicability and limitations of these techniques.