Macroscopic Internet Topology Data Kit (ITDK)
This page describes the historical ITDK release 2016-09, which consists of
- two related router-level topologies,
- router-to-AS assignments,
- geographic location of each router, and
- DNS lookups of all observed IP addresses.
This ITDK is produced from active measurements conducted on our Archipelago (Ark) measurement infrastructure. For the router-level topology, we used a subset of the IPv4 Routed /24 Topology Dataset, which is collected continuously. Specifically, we obtained the raw IPv4 topology by performing traceroutes to randomly-chosen destinations in each routed /24 BGP prefix using 132 Ark monitors located in 49 countries on Sep 4-18, 2016.
The two included router-level topologies are generated from the same IP-level topology but differ in the accuracy and completeness of the alias resolution performed to create them. The first topology is derived from aliases resolved with MIDAR and iffinder, which yield the highest confidence aliases with very low false positives. The second topology also uses MIDAR and iffinder but further includes aliases resolved with kapar, which significantly increases the coverage of aliases but at the cost of false positives (which inflate the size of routers and decrease the router count). Researchers should choose the topology to use depending on the relative importance they place on accuracy vs. comprehensiveness of alias resolution. Choose the most accurate alias resolution if uncertain about which to use.
Each router-level topology is provided in two files, one giving the nodes and another giving the links. There are additional files that assign ASes to each node, provide the geographic location of each node, and provide the DNS name of each observed interface.
The nodes file lists the set of interfaces that were inferred to be on each router.
node <node_id>: <i1> <i2> ... <in>
Each line indicates that a node node_id has interfaces i1 to in. Interface addresses in 22.214.171.124/3 (IANA reserved space for multicast) are not real addresses. They were artificially generated to identify potentially unique non-responding interfaces in traceroute paths.
NOTE: In ITDK release 2013-04 and earlier, we used addresses in 0.0.0.0/8 instead of 126.96.36.199/3 for these non-real addresses.
The links file lists the set of routers and router interfaces that were inferred to be sharing each link. Note that these are IP layer links, not physical cables or graph edges. More than two nodes can share the same IP link if the nodes are all connected to the same layer 2 switch (POS, ATM, Ethernet, etc).
link <link_id>: <N1>:i1 <N2>:i2
[<N3>:[i3]] .. [<Nm>:[im]]
Each line indicates that a link link_id connects nodes N1 to Nm. If it is known which router interface is connected to the link, then the interface address is given after the node ID separated by a colon (e.g., "N1:188.8.131.52"); otherwise, only the node ID is given (e.g., "N1").
By joining the node and link data, one can obtain the known and inferred interfaces of each router. Known interfaces actually appeared in some traceroute path. Inferred interfaces arise when we know that some router N1 connects to a known interface i2 of another router N2, but we never saw an actual interface on the former router. The interfaces on an IP link are typically assigned IP addresses from the same prefix, so we assume that router N1 must have an inferred interface from the same prefix as i2.
The node-AS file assigns an AS to each node found in the nodes file. We use our final Election+Degree assignment heuristic to infer the owner AS of each node.
node.AS <node_id> <AS> <method>
Each line indicates that the node node_id is owned/operated by the given AS, as inferred with the given method. There are three inference methods:
- a router has only a single choice of AS
- multiple ASes are present on a router, and one AS occurs more frequently than the rest
- multiple ASes are present on a router, but no AS occurs the most frequently, so the choice is based on AS degree
The hostnames file contains the hostname for every IP
address in the router-level topology for which a successful reverse
DNS lookup could be found.
<timestamp> <IP_address> <hostname>
The node-geolocation file contains the geographic location for each node in the nodes file. We use MaxMind's GeoLite City database for the geographic mapping.
- Data older than one year is available as a public dataset. You can obtain access using this form.
- The most recent one year of data is available for use by academic researchers and US government agencies. This data is also available for corporate entities (including corporate researchers) who participate in CAIDA's membership program. Please, complete and submit the online form to request access to the most recent data. It usually takes about two to three business days to process your request. Upon approval you will receive an email with instructions on how to download the data you requested. If you have any questions or problems using this form, please contact firstname.lastname@example.org.
Acceptable Use Agreement for the public data
Please read the terms of the CAIDA Acceptable Use Agreement (AUA) for Publicy Accessible Datasets below:
When referencing this data (as required by the AUA), please use:
The CAIDA Macroscopic Internet Topology Data Kit - <release dates >,
You are required to report your publications using this dataset to CAIDA.
Request Data Access
- Access the publicly available CAIDA Ark IPv4 Internet Topology Data Kits Dataset (and other topology data)
- Request Access to the restricted CAIDA Ark IPv4 Internet Topology Data Kits Dataset
Note that two historical ITDK releases made in 2002 and 2003 are also available as public datasets. These datasets should be used with caution, as they were constructed using completely different procedures and using topology data collected on the now decommissioned skitter measurement infrastructure.
Internet Topology Data Kit ProcessBelow we describe the various steps involved in producing the datasets that are part of the ITDK.
For alias resolution, we rely on several CAIDA tools: iffinder, kapar, and MIDAR, (recent tech report). MIDAR (Monotonic ID-based Alias Resolution, a tool we hope to release soon) expands on the IP velocity techniques of RadarGun, while kapar expands the analytical techniques of APAR. We use the traceroute dataset as input to MIDAR and iffinder, which generate output files used as input to kapar. kapar heuristically infers the set of interfaces that belong to the same router, and the set of two or more routers on the same IP link (a construct that represents either a point-to-point link, or LAN or cloud with multiple attached IP addresses).
We have an in-house bulk DNS lookup service called HostDB that can look up millions of addresses per day. We look up all intermediate addresses and responding destinations seen in the Topology Dataset. Each ITDK contains a list of the successful lookups for each IP address found in the nodes dataset.
To assign IP addresses to ASes, we used a publicly available BGP dump provided by Routeviews. BGP (Border Gateway Protocol) is the protocol for exchanging interdomain routing information among ASes in the Internet. A single origin AS typically announces ("originates") each routable prefix via BGP. We perform IP-to-AS mapping by assigning an IP address to the origin AS of the longest matching prefix for that IP address in the BGP tables.
The goal of the AS assignment process is to determine the AS that owns each router. For each router r, we first create an AS frequency matrix that counts the number of interfaces (known and inferred) from each AS that appears on r. The ASes in this frequency matrix represent the set of possible owner ASes of r. We use the following AS assignment heuristic to assign a router r to an AS.
The Election heuristic assigns router r to the AS with the highest frequency in r's AS frequency matrix. The intuition behind this heuristic is that routers will tend to have more interfaces in the address space of their owner. If two ASes from r's AS frequency matrix have the same count, then Election cannot decide an owner.
The Customer heuristic uses the AS relationship dataset to assign relationships to each pair of ASes from r's AS frequency matrix. Customer assigns r to the AS inferred to be a customer of every other AS in r's AS frequency matrix. This heuristic is based on the common practice that customer and provider routers typically interconnect using addresses from the provider's address space. Consequently, a router with interfaces from both the customer and provider address spaces is assigned to the customer.
For the Degree heuristic, we first generate an AS-level graph by assuming full-mesh connectivity among ASes from each router's AS frequency matrix. We then use this graph to generate an AS degree for each AS. Degree assigns router r to the smallest-degree AS from r's AS frequency matrix, i.e., the AS most likely to be the customer AS, based on similar intuition as the Customer heuristic.
For the Neighbor heuristic, we first determine the set of single-AS routers to which r is connected (its single-AS neighbors). We create a new AS frequency matrix that counts the number of single-AS neighbors of r from each AS. The Neighbor heuristic assigns r to the AS with the largest frequency (most single-AS neighbors), based on the intuition that a router is connected to a larger number of single-AS routers in its owner AS. Neighbor produces an ambiguous assignment when multiple ASes have the same (highest) frequency.
In case one of the previously described primary heuristics is unable to produce an AS assignment, we attempt to break the tie using one of the other heuristics as a tie-breaker. Our evaluation in the paper shows that Neighbor was the best stand-alone heuristic, while Election+Degree was the best combination.
For further details, please see the paper Toward Topology Dualism: Improving the Accuracy of AS Annotations for Routers.
We use MaxMind's free GeoLite City database to provide the geographic location (at city granularity) of routers in the router-level graph. Because this database maps individual IP addresses to locations, we take the following steps to find the location of each router (which by definition has multiple interfaces). We first map each interface on a router to a location. If all interfaces map to the same location, then we assign that location to the router; otherwise, we do not assign any location to the router (that is, the router does not appear in the geolocation file).