AT&T Labs - Research
AT&T  
Nick Duffield's Publications


Network Performance Tomography

Overview

Network Performance Tomography is the science of correlating performance measurements across a mesh of end to end path in order to infer link performance in the network interior. It provides algorithms to infer packet loss, delay, and even the underlying topology, from the end to end measurements. Our first paradigm for performance tomography exploited the inherent packet level correlations of multicast probes exchanged across a dedicated measurement infrastructure. Subsequently we have moved beyond inherent limitations of this approach by (i) emulating multicast probes with sets of unicast packets; (ii) exploiting simple performance level correlations amongst ordinary packet streams, and (iii) piggybacking probing and data collection onto ordinary network traffic and protocols and hosts to form a lightweight impromptu measurement infrastructure.

This work has been carried out partly in collaboration with the University of Massachusetts, ICIR, and the University of Paris.

Simple Network Tomography

Impromptu Measurement Infrastructures

Unicast Tomography

Multicast Tomography

Related Projects

Contact

Nick Duffield / duffield@research.att.com