Still under construction!

An Internet Topology for Simulation

Michael Liljenstam, Jason Liu, and David M. Nicol.

Background

Why do we need a real Internet topology for simulation?
  1. Because we can. By combining parallelism and modeling techniques, researchers now are able to efficiently simulate a network of very large size---with millions of network entities, under realistic network traffic conditions, in real time. The capability of simulating large network models creates the demand of using realistic network description of commensurate size.
  2. Synthetic network topology captures some characteristics of the global Internet at certain level, but cannot provide a full description of the network as it is today, especially missing information such as link bandwidths, latencies, and traffic routing behavior. That latter one is largely determined by economy reasons and business relations, which are highly dynamic. In fact, we believe there is no consensus on what are important attributes of network topology.
We try to capitalize on recent research on Internet mapping and measurement, and generate a realistic router-level network topology using the maps from those research results. In particular, we try to combine the data sets collected by multiple mapping tools, such as skitter from CAIDA, Mercator by the SCAN project from ISI, and RocketFuel from University of Washington. We seek to assign link attributes (i.e., bandwidths and delays) using information available from each ISP. Part of the information can be obtained from the PoP (Point-of-Presence) level map collected by the Mapnet project at CAIDA.

The Backbone Topology

We started from the RocketFuel data set, which contains router-level maps of several major ISPs. We chose six ISPs from RocketFuel. We added two more ISPs (UUNet and Cable & Wireless) from the SCAN data set. We believe the topology provides a good coverage of the Internet backbone in U.S. The following table shows the details of each ISP. The BGP routers are those that connect to another router belonging to a different ISP in this set. The result network is assigned with link bandwidths---using published information about link types between PoPs, and link delays---calculated from distance between routers.

ASNameTotal RoutersBGP RoutersBackbone RoutersBackbone BGP Routers
1239Sprintlink1053126646556
2914Verio649416684782
3356Level3150311848261
3561Cable & Wireless53992622232239
3967Exodus4324121132
6461Abovenet4895924439
7018AT&T1196110872946
701UUNet74142664259238
Total4422311919469793

The end result is a router-level U.S. network map with about 44,223 nodes and 68,681 (bi-directional) links. One way to cut down the size of the network is to consider only those routers at the backbone. RocketFuel labels each router with a rank, which is the number of hops to the routers that consist of the backbone. If we only consider those backbone routers, we end up with a network of 9,469 routers and 29,556 links. The network backbone is show in the following figure.

Local Area Network Models

We so far have in our collection three local arena networks:
  1. A synthetic network model. It's named campus network, containing 30 routers connecting 504 client hosts and 4 server hosts. This model has been used extensively by many researchers in their network simulation studies.
  2. A network model of a small university. This network topology we obtained contains 16 routers at the campus level and 90 dangling attachment points for connecting LANs within buildings. We attached local host clusters (of 43 hosts each) to these dangling points, resulting in a network of 3,886 routers and 469 links.
  3. A network model of a government national research lab. The network topology has 448 routers and 203 dangling attachment points where we attach local host clusters. The result is a network of 9,177 routers and 1,281 links.
All the network interface cards in these network models are assigned with calculated IP addresses to protect the real network. All these local area networks can be attached to the global network topology for a more detailed network model.

Data Format

The network topology is described in two different formats: one in XML, and the other in DML. The DTD definition of XML can be obtained here. Basically, XML defines a set of router and link attributes. Each router is assigned with an id and contains a list of network interfaces (with distinct IP address and bitrate). Each link is assigned with a bandwidth and contains a list of attachment points (i.e., network interfaces identified by the node id and the interface IP address). To allow efficient description of the network topology, an XML file can also import sub-networks described in other XML files. The IP prefix and the starting node id of the subnet must be provided together with the file name of the XML to be imported. In this way, all IP addresses in the sub-network are masked by the IP prefix and the ids of all routers in the sub-network are added with the starting node id, so that the IP addresses and the node ids are all unique for the entire network.

DML stands for Domain Modeling Language, which has been used extensively by the SSFNet project. DML is a recursively defined attribute list. The network model can be described easily with DML. Click here for an introduction of using DML to describe network models.

Download

Applications

There have been a number of network models used for simulations studies, which are based on the network topologies described in this page. We list some of them here:
The web page was prepared by Jason Liu. Last modified: Thu Dec 4 17:55:22 CST 2003