Infraestructura de red del proveedor de servicios IP

Initial topology discovery methodologies focused on BGP table dumps. The most obvious application is simply visualising the topology induced by the AS-paths they contain: the fact that an AS-path appears in a BGP table dump implies1that there is an inter-AS link (i.e. a BGP peering) between every consecutive pair of ASes in the path. An early example of AS- level topology visualisation is described in [61], which uses the general purpose topology visualisation tool Otter. However, BGP tables have also been used by a large body of work on topics as diverse as mapping IP addresses to AS numbers [87] and evaluating the quality of different topology generation approaches [128, 15]. A brief overview of this work follows.

Evaluating the impact of routing policies and BGP performance

BGP table data has been used to investigate the efficiency of the routing policies ordinarily used in the Internet and the efficiency of the BGP protocol itself. A comparison of routing table dumps with node-level topology [129] showed that in a significant number of cases the presence of routing policy constraints leads to the use of longer AS-paths than if BGP were free to choose the shortest path, and to longer node-level paths than would be used if the Internet used node-level shortest-path routing.

Inferring AS relationships

Pioneering work by Gao [48] used BGP table dumps to infer the commercial relationships between ISPs. Current BGP routing policies are complex and motivated not only by technical reasons but also by economic factors such as only carrying traffic for paying customers or controlling the cost of links [17, 64]; Gao simplifies relationships between ASes to three

1_{This is not guaranteed to be true, because BGP implementations make it possible to modify the attributes of} routes. This, for example, allows operators to insert into the AS-path the numbers of ASes which did not actually propagate the update. We are not aware of any efforts to evaluate the extent of the problem by the topology discovery community, possibly because of the feeling that other data quality and completeness issues have a much greater impact on accuracy than intentionally incorrect routing information. However, from an operational point of view the fear that BGP might be used for malicious purposes has spurred the development of various BGP security architectures such as as S-BGP [73] and soBGP [133].

3.4. METHODS USING BGP ROUTING DATA 23

classes, customer-provider (where one AS pays the other for “transit”, i.e. connectivity to parts of the Internet it cannot reach on its own), peer-to-peer (where two ASes exchange traffic between each other and each other’s customers without any financial compensation) and sibling-to-sibling (where two ASes provide mutual transit to each other, such as in the case of mergers and acquisitions between ISPs).

The basic assumption is that every AS providing transit must be paid for the service (and thus is a provider), and that every AS receiving transit pays for the service (and thus is a customer). Therefore, since IP traffic flows along the path taken by BGP announcements in the opposite direction, the commercial relationship between a given pair of ASes poses restrictions on the prefixes that may be announced on the peering between those ASes. For example, (i) a customer may announce to a provider its own prefixes and its customers’ prefixes, but not prefixes received from a peer or from a provider; (ii) a peer may announce to a peer only its own prefixes and its customers’ prefixes, but not prefixes received from a provider or from a peer; (iii) a provider may announce any prefix to a customer. These constraints can be applied to the AS-paths that appear in routing tables, and “valid” paths are paths that respect the constraints. Based on these constraints and the fact that a provider is typically larger than its peers and that peers are typically of comparable size, Gao proposes heuristics for inferring the types of inter-AS relationships and evaluates them by comparing them to internal AT&T information and IRR data, with very good results.

Further work on the problem was carried out by Subramanian et al. [127], who relaxed the problem by not inferring sibling-to-sibling links and provided a more elegant mathematical formulation in the form of an optimisation problem with the goal of maximising the number of valid paths. They named this problem the ToR problem, for Types of Relationships, and conjectured that it is NP-complete. They provided a heuristic solution and used it to give a five-level hierarchical classification of the ASes in the Internet, with a topmost level consist- ing of an almost complete clique between 20 “Tier 1” providers, which have no providers and peer amongst themselves. The NP-completeness of ToR was then proved by Erlebach et al. [44] and indipendently by Di Battista et al. [38], who also developed mathematically rigorous approximate solutions to the ToR problem and proved that peer-to-peer links cannot be inferred in the ToR problem formulation.

The quality of the inter-AS relationships inferred by these methods was analysed in [138], which compared their results to information on inter-AS relationships documented in IRR data and obtained by observing documented BGP community attributes [25]. The authors find that the accuracy of the relationships inferred is very high for customer–provider rela- tionships, but much lower (between 25% and 50%) for peer-to-peer relationships.

Evaluation of the quality of BGP information

BGP table dumps have been used to verify the quality of IRR data and vice versa. In [121], the authors assess the quality of data contained in the IRR system, showing that although IRR data can provide unique information that cannot be found in BGP table dumps, it also contains many inaccuracies. Conversely, Chang et al. [26] compares ORV table dumps to information collected from looking glasses and IRR data, showing that the table dumps only produce an incomplete picture of AS interconnections and that fusing the information from all data sources reveals a graph with 25-50% more AS interconnections. Finally, Broido et al. [14] compares BGP table dumps with node-level topology data, finding that the AS-level topologies induced by node-level data have much denser inter-AS connectivity than those observed in BGP table dumps.