InterNet Control and Inference Tools at the Edge

Network Tomography

[Simulated Network] [Bottleneck scenario]
Left: Tree structured network with a single-souce, multiple-receiver network. Node 0 is the source, nodes 1-4 internet routers, and nodes 5-11 receivers. Besides each link indicates the capacity in megabits per second.

Right: Network Simulator (ns-2) simulation result. True and estimated link-level success rates of TCP flows from source to receivers for heavy losses on links 1-2 and 2-5. (Bottom) Mean Absolute Error between estimated and true succes rates over 10 trials for each link.


In large-scale networks, end-systems cannot rely on the network itself to cooperate in characterizing its own behavior. It is also impossible to set up internal monitors at all routers in the Internet due to cost and deployment problems. In general, taking measurements at the edge of a network is a much easier and inexpensive proposition than performing internal monitoring. This has prompted us to study the so-called Network Tomography problem.


To present methodology for network tomography based on passive monitoring of unicast traffic. Parameters such as link loss rates, delay statistics and topology inference will be modeled and its performance will be evaluated with realistic network simulations (Network Simulator, ns-2).


|INCITE Home |Rice DSP |Rice University|